Brain-wide association studies try to link brain scans to behavior. Reproducible brain-wide association studies require thousands of individuals resampled roughly 50,000 people from three large cohorts (ABCD, HCP, UK Biobank) and found that the true correlations are tiny, with a median $|r|\approx 0.01$. At typical sample sizes, the reported effects are inflated by more than 200% because only noise spikes clear the significance threshold. The field's answer is federation: hospitals keep their scans and exchange only model updates. Platforms such as COINSTAC and MetisFL do exactly this, and they run into wide-area bandwidth and straggler problems (slow sites that hold up each round). The obvious networking fix is to borrow load-adaptive contention control from wireless and datacenter protocols. In these schemes each sender decides how often and how much to transmit from live signals such as its queue length or measured delay. The sender's "contention window" is the random waiting period it draws before trying to send.
This note argues that the obvious fix is statistically dangerous, and it proposes a version that is safe. If a scheduler decides which sites speak in a round from signals that correlate with the data, such as loss improvement, gradient size, or whether a site met a deadline, it performs a selection on outcomes. That selection can recreate the winner's curse at the level of the federation, and it does so at exactly the effect sizes (|r| of a few hundredths) where even small biases decide what gets reported. We propose a contention scheduler that adapts transmission probability and burst size only to outcome-blind signals: queuing delay, backlog, and site sample size. The aggregator records each site's inclusion probability and reweights by its inverse. This keeps the pooled estimate unbiased while gaining the throughput benefits of adaptive access.
Core Claim
Our claim has three parts. First, under a fixed wall-clock or byte budget, the participation policies people actually use in federated imaging inflate effect sizes and lower replication rates for small effects. These policies are synchronous rounds with straggler deadlines, utility-driven client selection, and backlog-proportional access in the style of O-DCF. The inflation is measurable relative to full-cohort ground truth. Second, the inflation comes from two sources: a site's chance of being heard depends on its data, and the aggregate is not corrected for that chance. Third, the following design matches the statistical behavior of full synchronous participation while recovering most of the wall-clock speedup of adaptive access. Inclusion probabilities are set by a sigmoid of an outcome-blind load signal and then corrected by inverse-probability weighting. Floors on those probabilities protect small sites.
The new element is not adaptive pacing as such. Prior systems already pace or schedule. The new element is the requirement that the pacing signal be ancillary to the estimand, meaning it carries no information about the quantity being estimated, together with the bookkeeping that makes the estimator design-unbiased.
Research Program
Source of the mechanism. We take the mechanism from Making 802.11 DCF near-optimal: Design, implementation, and evaluation. O-DCF maps a node's local queue backlog through a decreasing sigmoid to set its contention window, and it scales burst length with backlog. On a 16-node testbed this improved Jain's fairness by up to 87.1% over standard DCF. We also take the load signal from Accurate Latency-based Congestion Feedback for Datacenters. DX uses precise per-packet queuing delay rather than coarse congestion marks, and it cut switch queues by 4.85× relative to DCTCP at 1 Gbps.
Where the transplant breaks. The research file flags the tension (Gap 4) that O-DCF's backlog priority, carried over to a federation, would let a biobank site starve a 194-subject deeply phenotyped cohort such as A functional connectome phenotyping dataset including cognitive state and personality measures. Our contribution is to show that starvation is a statistical problem and not only a fairness one. When sites differ in population, which is the normal case across cohorts like cVEDA and IMAGEN, under-hearing a site shifts the estimand itself.
The scheduler. In round $t$, site $i$ transmits with probability
$$p_i(t) = \max\{\pi_{\min},\ \sigma(a\, b_i(t) - c\, d_i(t))\},$$
where:
- $b_i(t)$ is the site's normalized backlog of pending local computation;
- $d_i(t)$ is its measured queuing delay to the aggregator;
- $\sigma$ is the logistic sigmoid;
- $a$ and $c$ are gains;
- $\pi_{\min}$ is a floor that guarantees every site is heard.
Burst size, meaning how many local steps or how many matrix blocks are shipped, scales with $b_i$ as in O-DCF.
The aggregator. The aggregator forms
$$\hat{g}(t) = \sum_i \frac{\mathbb{1}[i \text{ sends}]}{p_i(t)}\, w_i\, u_i(t),$$
where:
- $u_i(t)$ is site $i$'s update or block of sufficient statistics;
- $w_i$ is its target weight, for example its share of the total $N$;
- the indicator equals 1 if site $i$ transmitted in round $t$.
This is the standard Horvitz–Thompson correction. It is unbiased for the full-participation aggregate exactly when $p_i(t)$ does not depend on $u_i(t)$ given the past. That condition is why the signals must be outcome-blind.
Analysis. We will bound the variance cost of the floor $\pi_{\min}$ against the bias that any outcome-dependent policy incurs. We will also characterize, in a simple two-site Gaussian model, when deadline-based dropping is biased. The answer we expect is that it is biased whenever site latency correlates with the site-specific effect.
Workloads. We study iterative analyses where rounds actually matter. These are the decentralized linear mixed-effects models of Decentralized Mixed Effects Modeling in COINSTAC, the multi-shot dMANCOVApa of Federated analysis in COINSTAC reveals functional network connectivity and spectral links to smoking and alcohol consumption in nearly 2,000 adolescent brains, and 3D CNN training as in Secure Federated Learning for Neuroimaging. In the CNN case encrypted parameters roughly double message size, so bandwidth pressure is real.
For neural models we use FedAdam from Adaptive Federated Optimization as the server optimizer. That paper treats rounds abstractly and ignores the channel, so our scheduler sits beneath it without changing it.
Baselines
Reference point. The naive reference is synchronous full participation with no deadline: every site is waited for in every round. This defines the statistical target that every other policy should reproduce. It also sets the wall-clock cost that the other policies try to reduce.
Practical default. The baseline most deployments actually use is synchronous rounds with a fixed timeout, where stragglers are dropped. This is the static round scheduling that COINSTAC and MetisFL assume. We run it with the same timeout budget as the proposal's wall-clock target.
Strongest throughput competitor. This is O-DCF-style backlog-proportional access without inverse-probability correction, a direct transplant of Making 802.11 DCF near-optimal: Design, implementation, and evaluation. It should match our speed. If it also matched our bias and coverage, the correction would be unnecessary.
Utility-driven selection. We include a policy that favors sites whose last update most reduced global loss or had the largest norm. This represents the outcome-dependent selection we predict to be harmful. It is common in federated learning practice, though no instance appears in the research file.
Ablations. Two ablations isolate the key ingredients. The first removes the floor $\pi_{\min}$ while keeping reweighting, which tests whether small-site variance blows up. The second keeps reweighting but uses an outcome-dependent signal, which tests whether reweighting alone saves a non-ancillary scheduler. It should not, because $p_i$ then depends on $u_i$.
Fairness controls. All policies share the same simulated network traces, site partitions, local computation, server optimizer, and hyperparameter search budget. Each is tuned on held-out partitions and never on the evaluation partitions.
What counts as success. For the 3D CNN Alzheimer's task, the strongest reported reference is MetisFL's federated AUC of 0.8756, against a centralized AUC of 0.8789 (Secure Federated Learning for Neuroimaging). Beating the baselines means two things. The proposal must stay within that federated-versus-centralized gap at a lower wall-clock cost than synchronous full participation. It must also not lose small-site per-site AUC, which in that paper falls as low as 0.53 for isolated sites.
For the server optimizer, there is no reported result on Lacuna for FedAdam on neuroimaging. Adaptive Federated Optimization reports only qualitative superiority over FedAvg on its own benchmarks. For the association tasks, success means two things relative to the deadline and uncorrected-O-DCF baselines. Effect-size inflation must be statistically indistinguishable from full participation. Confidence-interval coverage must be at nominal level.
Why not just…?
Why not just ship one-shot sufficient statistics and skip scheduling entirely? For univariate or ordinary least-squares brain-wide association, each site can send $X^\top X$ and $X^\top y$ once, and the pooled result is exact. This is a real threat. We concede it: for those analyses we use one-shot pooling as ground truth rather than as a competitor. The proposal is scoped to analyses that are multi-shot by necessity. Mixed-effects models with site random effects iterate in Decentralized Mixed Effects Modeling in COINSTAC, and CNN training iterates by construction. Where one-shot pooling suffices, we expect and will report no benefit.
Why not just wait for every site, as plain synchronous FedAvg does? Waiting for everyone is unbiased, which is why it is our reference. Its cost is wall-clock time set by the slowest, encrypted, wide-area link. The experiment measures how much of that time the proposal saves at equal bias. If the saving is small on realistic traces, the honest conclusion is that clinical federations should simply wait, and we will report that.
Why not just let TCP handle congestion and drop sites that miss the deadline? Transport-level control keeps links stable, but it says nothing about which sites' statistics end up in the estimate. Deadline dropping is outcome-blind only if lateness is unrelated to site populations. Large sites, which are slow, often differ demographically from small ones. Gazula et al. federate an Indian and a European cohort (Federated analysis in COINSTAC reveals functional network connectivity and spectral links to smoking and alcohol consumption in nearly 2,000 adolescent brains), where lateness and population plausibly co-vary. Our size-skewed, population-shifted partitions test this directly. If deadline dropping shows no measurable bias there, the proposal's statistical motivation weakens to a pure speed argument.
Why not just go asynchronous with staleness down-weighting? Asynchronous updates arrive at rates set by site speed. Without inclusion-probability bookkeeping, fast sites are over-represented in exactly the way uncorrected O-DCF over-represents them. We include asynchronous FedAvg as an additional uncorrected arm under the same traces. We expect its bias to resemble that of uncorrected O-DCF.
Proposed Experiments
Association experiment. The decisive experiment is a resampling study in the style of Reproducible brain-wide association studies require thousands of individuals, run on ABCD and UK Biobank functional connectivity and behavior.
- Ground truth. We take full-cohort estimates as the reference.
- Federations. We repeatedly draw federations of 5 to 20 virtual sites. Their size distributions are skewed, ranging from tens to thousands of subjects, and some partitions carry deliberate population shifts across sites.
- Network conditions. We assign each site a simulated wide-area link with heterogeneous bandwidth and bursty delay. Message sizes reflect plaintext and encryption-doubled payloads.
- Model. We run a decentralized mixed-effects association model under each policy at matched wall-clock budgets.
- Metrics. For significant edges, we measure inflation of the estimated effect relative to the full-cohort truth. We also measure 95% interval coverage, the replication rate between disjoint federations, small-site share of effective weight, and bytes and wall-clock time per converged fit.
The prediction is specific. The deadline, utility-driven, and uncorrected O-DCF policies will show inflation and coverage loss that grow with population shift. The proposal will track full participation.
Prediction task. The second experiment repeats the comparison for Alzheimer's classification across ADNI, OASIS-3 and AIBL with the MetisFL setup, using encryption-sized messages. We measure global AUC against 0.8756 and 0.8789, per-site AUC for the smallest site, and wall-clock time.
Ancillarity check. A third check confirms that the ancillarity condition is the operative ingredient. Only the arm that combines inverse-probability weighting with an outcome-dependent signal should fail coverage. That outcome would show that reweighting is necessary but not sufficient, and that the signal choice carries the guarantee.
References
- Reproducible brain-wide association studies require thousands of individuals
- Making 802.11 DCF near-optimal: Design, implementation, and evaluation
- Accurate Latency-based Congestion Feedback for Datacenters
- A functional connectome phenotyping dataset including cognitive state and personality measures
- Decentralized Mixed Effects Modeling in COINSTAC
- Federated Analysis in COINSTAC Reveals Functional Network Connectivity and Spectral Links to Smoking and Alcohol Consumption in Nearly 2,000 Adolescent Brains
- Secure Federated Learning for Neuroimaging
- Adaptive Federated Optimization