Children with long COVID (the lingering illness after a SARS-CoV-2 infection, called post-acute sequelae or PASC) are now mostly characterized by mining hospital records for diagnoses that cluster together. Nearly all of that mining happens inside large academic children's hospitals. Children seen there are seen often, by many specialists, so any condition they have is more likely to be written down. In network science this failure mode is well understood. If you observe a network only through a narrow local window, where nearby links are over-represented, you see structure that is not there: inflated clustering, spurious reciprocity, false "small-world" wiring. The standard fix is to simulate the sampling process itself and ask what structure it would create on its own. This proposal applies that fix to pediatric PASC. A child's own record is a local observation window whose width is set by how often the child is seen. Because every diagnosis a child has is recorded through that same window, diagnoses will appear to co-occur even when they are biologically unrelated.
The concrete contribution is a sampling-matched null model for pediatric phenotype discovery. We estimate each child's observation intensity from the pre-infection "washout" period, which PASC cannot have caused. We then use that intensity to generate records that match the real data in how often children are seen but contain no real post-viral coupling. Finally, we run the published discovery pipelines on those null records. The residual, meaning the structure that survives this null, is the part of the pediatric PASC phenotype map we should believe.
Core Claim
The claim is that part of the multisystem structure reported for pediatric PASC is an observation-intensity artifact, and that this part can be measured and removed. A simple recording model shows why. Suppose child $i$ has $n_i$ clinical encounters in the post-acute window, and a true condition $d$ is coded at each encounter with probability $p_d$. The chance the condition appears in the record is then
$$P(\text{coded}_{id} \mid \text{has}_{id}) = 1 - (1 - p_d)^{n_i}.$$
Every condition a child has passes through the same $n_i$. Two conditions that are statistically independent in biology therefore become positively correlated in the record whenever $n_i$ varies across children. Tertiary-care cohorts are exactly where $n_i$ varies most: 55.8% of PASC cases in Lorman et al. had pre-existing chronic conditions. This is the clinical analogue of distance-dependent sampling.
The obvious fix is to adjust for the number of post-acute encounters, and it is wrong. PASC itself generates encounters, so conditioning on post-acute $n_i$ also removes real signal. What is new here is using pre-infection intensity, measured in the 18-month washout window that Lorman et al. already define, as an estimate of observation breadth that the infection cannot have caused. We then propagate that estimate through a generative null rather than a regression adjustment.
Relation to Prior Work
The mechanism comes from network neuroscience. Miner and Triesch showed in simulated cortex that restricting observation to a spatial window inflates connection density and reciprocity. Gerhard et al. showed that compact electrode arrays make recorded networks look small-world. She, Chen and Chan showed that efficiency metrics inflate as the sample shrinks, and argued that the sampling process must be modeled explicitly. From these papers we reuse two techniques: forward-simulating the observation process on a structureless null, and tracking how a metric changes as the observation window widens. Yang, Ribeiro and Neville motivate the clinical framing. They found that crawls trapped in dense neighborhoods bias effect estimates, and a child's referral path from clinic to specialist is such a crawl.
The object of study is three RECOVER analyses. Lorman et al. used a tree-based scan over ICD-10 codes (the diagnosis coding system) and found 304 significant diagnostic clusters. Rao et al. used a site-adaptive latent class analysis, a model that sorts patients into unobserved groups, and found three MIS-C severity classes (47.9%, 23.3%, 28.8%). Zhang et al. used difference-in-differences with negative control outcomes to estimate racial and ethnic PASC risk ratios.
All three adjust for measured confounders or for site. None models how often each child is observed as a mechanism that generates co-occurrence. Zhang et al. explicitly leave ascertainment bias unresolved. The adult subphenotypes of Zhang et al. (Poisson factor analysis) serve as a secondary pipeline to stress-test.
Two network methods in the file constrain the design rather than supply it. Network Kriging needs a known linear measurement matrix and some fully measured paths to calibrate against. We have neither for children who never reach a PEDSnet hospital, so we do not claim to reconstruct community prevalence. The mean-field recovery of Jiang, Gao and Magdon-Ismail assumes known ordinary differential equations (ODEs) governing the system, which clinical codes do not provide. We set both aside deliberately.
Research Program
The first component is the observation model. For each child in RECOVER PEDSnet we fit a hierarchical recording model on the pre-infection washout period. It estimates per-diagnosis coding probabilities $p_d$ and a per-child intensity trajectory. The model is stratified by site, age, chronic-condition burden and care setting (outpatient, emergency department, inpatient). We validate it by predicting pre-infection co-occurrence in held-out COVID-negative children. If intensity alone reproduces most pre-infection co-occurrence among conditions believed to be unrelated, the artifact is real before PASC enters the picture.
The second component is the null generator. For each real PASC case and control, we sample a synthetic record with three properties. First, conditions are drawn independently from the child's own marginal rates. Second, the post-acute encounter count is drawn from the washout-predicted intensity. Third, codes are recorded by the equation above. The same case definition is then applied to the synthetic cohort. This step matters because PASC case status itself requires contact with the health system, which makes the selection differential between cases and controls.
The third component re-runs the published pipelines on the synthetic cohorts: the tree-based scan, latent class analysis, and Poisson factor analysis. We record the structure metrics each pipeline reports. These are the number of significant cuts, the median number of organ systems per class, class separation (entropy), and the clustering coefficient of the organ-system co-occurrence graph. A finding counts as supported only if its real-data value exceeds the upper tail of the null distribution.
The fourth component is a window-width analysis in the style of She et al.. We stratify children by pre-infection intensity and plot each structure metric against intensity. Real post-viral coupling should be roughly flat across strata. Artifact should rise with intensity.
Baselines
The naive reference point is the published analyses run as-is on the same PEDSnet extract. For Lorman et al. that means 304 significant cuts. For Rao et al. it means the three severity classes. For Zhang et al. it means the reported risk ratios, including POTS/dysautonomia RR 1.57 in Asian American and Pacific Islander children and cognitive symptoms RR 1.25 in non-Hispanic Black children. We reproduce these numbers first. A reproduction that diverges invalidates later comparisons.
The strongest existing method is the negative-control-outcome calibrated difference-in-differences design of Zhang et al.. It is the only published pediatric approach that explicitly targets healthcare-seeking bias. We run it on the identical cohort with its published negative control outcomes, and we give it the same freedom to choose negative controls that our model gets to tune strata. No reported result was found on Lacuna for any method's accuracy at recovering known ground-truth phenotype structure under ascertainment bias. Its reported risk ratios are therefore the comparison values, and the semi-synthetic benchmark below supplies accuracy.
A second strong competitor is inverse-probability weighting by post-acute encounter count, together with the site-adaptive latent class adjustment of Rao et al..
The key ablation isolates the ingredient we claim matters. It is identical to our method except that intensity is estimated from post-acute rather than pre-infection encounters. If the two versions perform the same, the washout idea adds nothing.
Our method beats the baselines if it meets two conditions on the semi-synthetic benchmark. First, it must recover planted structure with higher adjusted Rand index and cut-level precision than both the negative-control design and post-acute weighting. Second, it must do so without the ablation's loss of recall on true planted effects.
Why not just…?
Why not just adjust for the number of encounters? Encounters after infection are partly caused by PASC, so the adjustment conditions on something downstream of the exposure and attenuates genuine effects. This is the ablation above. We expect it to match our method on precision and lose on recall of planted true effects. If it does not lose recall, this alternative is sufficient and we will say so.
Why not just rely on negative control outcome calibration? Negative controls remove bias that shifts all outcomes equally, such as a general tendency to seek care. The artifact here is multiplicative and differs by condition, because rarely coded conditions gain most from extra encounters. Zhang et al. themselves list ascertainment as unresolved. This alternative is a real threat for the risk-ratio analyses, and the semi-synthetic benchmark, which plants effects of differing coding rarity, will show where it succeeds.
Why not just restrict to children with regular well-child visits? This is the "widest window" stratum of our intensity analysis and is a legitimate sensitivity check. It trades bias for sample size and changes the population. With about 1,309 cases in Lorman et al., restriction may leave too few cases for rarer clusters. We report it as a separate estimate rather than a replacement.
Why not reconstruct community-level PASC with network kriging? Network Kriging achieved 10% relative error from 7 of 30 paths on Abilene. It did so because the routing matrix was known and calibration paths were measured. No pediatric data in the file gives either, so kriging here would be extrapolation without an anchor.
Proposed Experiments
The first experiment is a semi-synthetic benchmark, which also serves as the primary accuracy test. We start from real COVID-negative PEDSnet records. Into them we plant known post-acute patterns: independent conditions, and correlated multisystem modules with specified strength and coding rarity. We then impose recording by the fitted intensity model and apply each method. The metrics are adjusted Rand index against planted subphenotypes, precision and recall of planted tree cuts, and bias in planted risk ratios.
The second experiment is the real-data audit. We report, for each published finding, where its value falls in the null distribution. This covers the 304 cuts (with particular attention to the dominant symptom branch R00–R99), the MIS-C classes, and the subgroup risk ratios. The prediction we will test is that multisystem breadth and co-occurrence clustering fall substantially under the null, while the strongest single-condition signals such as dyspnea and fatigue survive.
The third experiment is the decisive check, the intensity dose-response. A genuinely post-viral cluster should show similar effect size in the lowest and highest pre-infection intensity strata. If most of the 304 cuts or the severe multisystem class grow steeply with intensity and vanish in the lowest stratum, pediatric PASC phenotyping in tertiary cohorts needs this correction as standard practice. If instead structure is flat across strata and survives the null, the published maps are robust, and that result is also worth publishing.
Replication across the 13-system RECOVER cohort of Zhang et al., whose care settings are more varied than PEDSnet's, tests whether the fitted intensity model transfers across networks. All analyses require RECOVER data-use approval. Code will be released alongside the RECOVER analysis environment.
References
- Understanding pediatric long COVID using a tree-based scan statistic approach: An EHR-based cohort study from the RECOVER Program
- Slicing, sampling, and distance-dependent effects affect network measures in simulated cortical circuit structures
- Extraction of Network Topology From Multi-Electrode Recordings: Is there a Small-World Effect?
- Evaluating the Small-World-Ness of a Sampled Network: Functional Connectivity of Entorhinal-Hippocampal Circuitry
- Should We Be Confident in Peer Effects Estimated From Social Network Crawls?
- Spectrum of severity of multisystem inflammatory syndrome in children: an EHR-based cohort study from the RECOVER program
- Racial/ethnic differences in post-acute sequelae of SARS-CoV-2 in children and adolescents in the United States
- Data-driven identification of post-acute SARS-CoV-2 infection subphenotypes
- Network Kriging
- True Nonlinear Dynamics from Incomplete Networks