When a child has lingering symptoms after COVID-19, several people can describe them: the child, a parent, and a clinician who may or may not record a diagnosis code. These accounts often disagree, and the disagreement is systematic, not random. Teenagers tend to hide symptoms that feel shameful or that might bring adult scrutiny. Anxious parents may over-report. Clinicians rarely code vague complaints such as fatigue or "brain fog" at all. Pediatric long COVID research currently picks one of these voices and ignores the rest. The consequences show up as contradictory numbers. A Beijing survey of 2,001 children found 55% with persistent symptoms. In a nine-hospital US health-record cohort, only 1,309 children had a formal post-COVID code.
This proposal borrows one idea from research on dividing authorship credit and applies it to symptom reports. That literature showed that giving every co-author full credit for a paper inflates totals, because the same contribution gets counted many times. Its fix is to treat credit as a fixed amount that is split among contributors. We argue the same inflation occurs when a parent and child who talk to each other both report "fatigue" and a model counts that as two independent pieces of evidence. We propose three things. Each symptom in each child carries one fixed unit of evidence. That unit is split among informants according to who can actually observe that symptom. Each informant's report is first corrected for a desirability shift, which we measure directly by randomizing the setting in which the child answers, instead of inferring it from a personality scale.
Core Claim
Multi-informant symptom reporting in pediatric post-acute sequelae of SARS-CoV-2 (PASC, or long COVID) has two separable distortions, and current practice corrects neither. The first is double counting: parent and child reports are correlated, so adding or averaging them overstates the evidence. The second is directional social desirability bias, meaning people shade answers toward what seems acceptable. This bias pushes in different directions for different symptoms and different informants. Our claim is that a fused phenotype built from conserved, symptom-specific informant shares, with each report corrected by an experimentally identified desirability shift, will predict later objective health-record events for subjective symptoms better than three alternatives: health-record codes alone, naive averaging, and a latent-class rater model that assumes informants err independently. We also claim it will do so without inflating symptom rates in COVID-negative children.
Grounding in Prior Work
The motivation comes from the PASC literature. Lorman et al. ran a tree-based scan over diagnosis codes in RECOVER PEDSnet, comparing 1,309 PASC cases with 6,545 matched COVID-positive and 6,545 matched COVID-negative controls. Their strongest signal was in the nonspecific "symptoms and signs" branch: dyspnea, chest pain, and fatigue. They also note that un-coded symptomatic children leak into control groups and bias estimates toward no effect. Zang et al. quantify the cost of relying on structured records. Their models reach a C-index above 0.8 for dementia, heart failure, and kidney failure, but only about 0.6 for fatigue, sleep disorders, and anxiety. The C-index measures how often a model ranks a person who develops the outcome above one who does not, so 0.5 is chance. Reese et al. found six reproducible PASC subtypes from Human Phenotype Ontology terms, and they too flag brain fog as under-recorded. Li et al. sit at the other extreme. They report high survey prevalence, including concentration difficulty in 22.1% of children, but treat guardian reports and older-child self-reports as interchangeable and name recall and subjectivity as limitations. Together these papers motivate the work: the subjective symptoms that define pediatric PASC are exactly the ones no single informant measures well. None of them combines informants.
The social desirability literature constrains the design. Studies of adolescents' reports of smoking and vaping under peer norms (Murray et al.), and of self-harm and bullying (Hou et al.; Zhang et al.), show suppression of stigmatized content. Work on health-behavior surveys (Tian et al.) shows inflation of socially praised behavior. The bias therefore changes sign across content, which rules out a single global correction factor.
The method comes from contribution weighting. Soen et al. normalize each citation to one unit of influence split across contributors. Nishioka et al. show that fractional splitting removes distortions in institution-level aggregation. Arandjelović and Wu's w-index instead fix shares by author position. Arandjelović also notes that position fails when it does not track contribution, for example when the senior investigator is listed last. That failure is the reason we do not assign informant shares by a fixed rank. Hu et al. tune a mixing weight between rank-decay and content signals ($\lambda = 0.56$) and validate against expert judgment. We take from them the practice of learning shares against an external target instead of declaring them. We differ from all of these in two ways: our "contributors" are informants rather than authors, and their shares vary by symptom.
Research Program
The estimator has two layers. Let $y_{ijs}$ be informant $j$'s report of symptom $s$ for child $i$, where $j$ is the child, the parent, or the clinician record. The first layer removes the desirability shift $\delta_{js}(a)$, which is specific to informant, symptom, and age band $a$. The second layer combines the corrected reports with shares that sum to one:
$$E_{is} = \sum_j w_{js}(a)\,\big(y_{ijs} - \delta_{js}(a)\big), \qquad \sum_j w_{js}(a) = 1,\; w_{js}\ge 0.$$
The constraint that the shares sum to one is the conservation idea taken from bibliometrics. Agreement between a parent and child cannot push $E_{is}$ above what one fully trusted informant would provide. The shares are therefore forced to express relative trust, not accumulated volume. We learn the shares $w_{js}(a)$ by maximizing how well $E_{is}$ predicts held-out objective anchors in the following six months. These anchors are new diagnosis codes in the relevant branch of the code hierarchy, specialist referrals, and encounter counts. We regularize the shares toward a prior based on who can observe the symptom: the child for internal sensations such as headache, fatigue, and concentration; the parent for observable behavior such as sleep timing and school avoidance; the clinician for signs such as hematuria or abnormal laboratory values.
The desirability shifts are identified by design, not modeling. Within each family we randomize whether the adolescent completes the symptom instrument privately on a tablet or with the parent present. We also randomize whether the parent is told the child's answers will be shared. Because assignment is random, the difference in mean report between settings estimates $\delta_{js}(a)$ for each symptom. This is the step that a uniform personality-scale correction cannot perform.
The data plan is a linked substudy. No dataset in our source material links child self-report, parent proxy report, and health records, so this linkage must be built. We would recruit COVID-positive and COVID-negative families at a RECOVER PEDSnet site, with children aged 8–17 and an age-appropriate instrument for 8–11-year-olds. We would link their reports to the site's records using the same 18-month washout and propensity matching as Lorman et al., then follow them for six months. Before fieldwork, we run a simulation. It is calibrated to the survey-versus-coding gap described above, with known correlation between informants and known desirability shifts that change sign. The simulation checks whether the estimator recovers the true shares and shifts at the planned sample size.
Baselines
The naive reference is the health-record-only approach. Each subjective symptom is predicted from structured codes and demographics with the gradient-boosted and regularized Cox models of Zang et al., re-fit on our linked cohort. For fatigue, sleep disorders, and anxiety they report a C-index of about 0.6. That figure is from adult and all-age cohorts, not children, so we also re-estimate it on our pediatric data before comparing.
The second baseline is equal-weight fusion: the mean of parent and child reports, plus a union rule in which any informant's report counts. This is what most survey studies implicitly do, including Li et al..
The third is fixed rank weighting, with shares of 1, 1/2, and 1/3 for clinician, parent, and child, in the style of Arandjelović.
The fourth is post-hoc deflation: a social desirability inventory administered alongside the instrument, with its score linearly subtracted from all symptom reports.
The strongest competitor is a latent-class rater model in the style of Dawid-Skene. It treats each child's true symptom status as hidden and learns each informant's sensitivity and specificity per symptom. Asymmetric sensitivity and specificity can already represent directional bias, which is why it is the main threat. No reported result for such models in pediatric PASC was found on Lacuna.
All baselines use identical linked data, the same anchor outcomes, the same train and test split by site-month, and the same hyperparameter search budget. We run three ablations that isolate the key ingredients. One removes the sum-to-one constraint. One replaces the randomized desirability shifts with shifts inferred by the latent-class model. One replaces symptom-specific shares with a single share per informant.
Beating the baselines requires two things. On the subjective-symptom anchors, the pooled C-index must improve over the latent-class model, with a confidence interval excluding zero. The method must also not raise the estimated symptom rate in COVID-negative controls relative to the health-record baseline by more than the simulation's tolerance.
Why not just…?
Why not just use diagnosis codes and drop questionnaires? The 0.6 C-index reported by Zang et al. and the control-group leakage described by Lorman et al. indicate that codes miss the very symptoms in question. Our first baseline tests directly whether adding informants helps at all. If it does not, the proposal fails at its first step.
Why not just average parent and child? Averaging treats correlated reports as independent and applies the same weight to every symptom, even though the bias changes sign across symptoms (Hou et al. versus Tian et al.). If equal-weight fusion matches our method on the anchors, the symptom-specific shares are not doing work, and we would report that.
Why not just subtract a social desirability scale score? A single scalar cannot represent suppression of concentration problems alongside inflation of healthy sleep habits in the same child. The randomized reporting-setting estimates test this directly: if the estimated shifts $\delta_{js}$ all share one sign and roughly one size, the scale correction is adequate and our design is unnecessary.
Why not just fit a latent-class rater model? This is a real threat. Such a model already learns informant reliability by symptom. Our method differs in two ingredients: it stops dependent informants from being double-counted, and it measures desirability shifts experimentally. The two ablations settle which ingredient matters. If a latent-class model that also allows parent-child dependence matches us once it is given our randomized shift estimates, then the contribution lies in the reporting-setting experiment rather than in conserved weighting, and we would present it that way.
Decisive Checks
Three results would decide the claim. The first is the comparison on held-out subjective-symptom anchors between our method and the latent-class model, with and without the sum-to-one constraint. The second is a placebo test on outcomes no one expects to follow COVID, such as injury-related codes. The fused phenotype should show no difference there between COVID-positive and COVID-negative children, which guards against gains that come only from inflating anxious-parent reports in exposed families. The third is the sign pattern of the randomized desirability shifts. The account predicts that adolescents report less concentration difficulty and fatigue when the parent is present, and that the shift differs between 8–11 and 12–17. If the shifts are null, the desirability half of the proposal should be dropped and only conserved fusion retained.
References
- Understanding pediatric long COVID using a tree-based scan statistic approach: An EHR-based cohort study from the RECOVER Program
- Identification of risk factors of Long COVID and predictive modeling in the RECOVER EHR cohorts
- Generalizable Long COVID Subtypes: Findings from the NIH N3C and RECOVER Programs
- Clinical manifestations and long-term symptoms associated with SARS-CoV-2 omicron infection in children aged 0–17 years in Beijing: a single-center study
- Moderators of peer influence effects for adolescents’ smoking and vaping norms and outcomes in high and middle-income settings
- Clustering analysis of risk factors for non-suicidal self-injury behaviors in adolescents: a cross-sectional study of western China
- Association between school bullying and self-harm in Chinese children and adolescents: the mediating role of mindfulness
- Pathways of influence on health-promoting lifestyles in older adults with coronary heart disease: a cross-sectional study
- Influence Flowers of Academic Entities
- How Does Author Affiliation Affect Preprint Citation Count? Analyzing Citation Bias at the Institution and Country Level
- Fairer Citation Based Metrics
- W-Index: A Weighted Index for Evaluating Research Impact
- A Novel Scholar Embedding Model for Interdisciplinary Collaboration