When a child stays sick for months after an infection, a parent loses work hours. How many hours a parent loses depends on the parent's job. A consultant can take a video call from a pediatric clinic waiting room. A nurse on a fixed shift, or a line worker, cannot. Economic models of epidemics that trace damage industry by industry count only the working adults who are themselves acutely ill, hospitalized, dead, or locked down. They treat children as carriers of infection and nothing more. Pediatric medicine, meanwhile, now documents long-lasting, multi-organ complications after infection in children, known as post-acute sequelae or "long COVID." It has never asked what those complications cost the adult workforce, or which industries absorb that cost.
We propose to measure the care burden pediatric long COVID actually creates, using the timing and type of children's clinical encounters in electronic health records. We would then convert that burden into a monthly, industry-specific loss of labor supply for the children's parents. This loss would be fed into an established input-output model, which tracks how a shortfall in one industry spreads to the industries that depend on it. The central question is whether this caregiver channel damages a different set of industries than acute adult illness does. If the ranking changes, pandemic economic planning is misallocating attention. If it does not, the channel can be folded into existing models as a simple multiplier. Either answer is worth knowing.
Core Claim
The closest existing work is the coupled epidemic–economy model of Epidemiological and economic impact of COVID-19 in the US. It links an age-structured infection simulation over a synthetic US contact network to a Dynamic Inoperability Input-Output Model. That model tracks, for each of 21 industries, the fraction of normal output lost, and how that loss spreads through supply dependencies. It finds that high-value industries such as Professional Services lose the most in dollars, even after accounting for how much of their work can be done remotely. The model assumes that survivors return to full productivity and that children matter only for transmission. The other near neighbor, COVID-EENet, predicts district-level spending by business category over 14-day windows. It says nothing about labor or about horizons of months.
Our claim is specific. Pediatric post-acute care demand arrives as a long, recurring stream: scheduled specialist visits, unscheduled emergency visits, and days the child cannot attend school. Remote work helps only with some of this. The parts that are fixed in time, such as a 10 a.m. cardiology appointment or a school pickup after a fever, cannot be absorbed by jobs with rigid schedules, whether or not the job can be done from home. We therefore predict that healthcare, manufacturing, retail, and transportation move up the vulnerability ranking under the caregiver channel, relative to their position under acute worker illness. We also predict a feedback loop specific to healthcare. That sector both supplies pediatric care and employs many parents who lose hours to consuming it.
Research Program
The first component turns children's health records into caregiver time demand. We work in RECOVER PEDSnet, which contains records from nine US children's hospitals and applies an 18-month look-back window so that new conditions can be separated from existing ones. We reuse the cohort design of the tree-based scan study of pediatric long COVID. That study compared 1,309 children with long COVID against 6,545 infected controls and 6,545 uninfected controls. It scanned the full hierarchy of diagnosis codes and found 304 significantly enriched diagnostic groups, concentrated in non-specific symptoms such as dyspnea, chest pain, and fatigue.
That study asked which diagnoses are enriched. We ask how much parental time the enriched diagnoses consume, and when. For each child, we extract the excess encounter stream relative to matched controls, by month from 1 to 12 after infection. Each encounter is split into three demand types. Scheduled outpatient visits are known in advance. Unscheduled emergency or urgent visits cannot be planned. Recurring treatment series, such as physical therapy or diabetes follow-up, repeat on a fixed calendar. Each type is converted into parent hours using setting-specific durations plus travel time, and the result is a demand curve $D_k(t)$ for type $k$ in month $t$.
Encounters are grouped by the diagnostic clusters from the scan study, so that different symptom patterns can have different demand shapes. The temporal findings of the distributed-representation study of long COVID sub-phenotypes show that some symptoms appear late: neurological deficits at a mean of 289 days and fatigue and dyspnea at 330 days, in hospitalized adults. For that reason we extend follow-up past 180 days wherever the data allow.
The comparison against infected controls matters for a reason beyond epidemiology. The type 1 diabetes study in RECOVER found a non-significant rise in diabetic emergencies after infection (hazard ratio 1.45, 95% CI 0.97–2.16). Blood-sugar control first worsened and then improved beyond day 180. That pattern raises the possibility that extra monitoring after infection, rather than disease, generates the extra visits. For clinical science this is a confound. For labor economics it is not: a surveillance visit costs a parent the same hours as a disease-driven one. We therefore report two quantities. One is the total excess demand relative to uninfected children, which is the economically relevant burden. The other is the portion attributable to infection itself, estimated against infected children who did not develop long COVID.
The second component maps that demand onto industries. For industry $j$, the monthly fraction of labor lost through this channel is
$$q_j(t) = \frac{\rho_j(t)}{H}\sum_k \big(1 - a_{jk}\big)\, D_k(t).$$
Here $\rho_j(t)$ is the share of industry $j$'s workers who are parents of a child with post-acute sequelae in month $t$. It is computed from pediatric attack rates by age combined with the distribution of parents' industries by child age. $H$ is contracted monthly hours. $a_{jk}$ is the fraction of demand type $k$ that industry $j$ can absorb without lost hours.
For demand that can be handled from any location, we set $a_{jk}$ using the 21 teleworkability fractions from Epidemiological and economic impact of COVID-19 in the US. For demand fixed in time, $a_{jk}$ must come from a separate measure of schedule rigidity. Two of the inputs, parents' industry by child age and schedule rigidity by industry, come from national household and labor-force surveys. The research file does not cover these sources, and we flag them as external dependencies. Because $a_{jk}$ for time-fixed demand is uncertain, we sweep it over a bounded range and report which conclusions hold across the whole range.
The third component feeds $q_j(t)$ into the input-output model as a persistent labor disruption spanning months. The model then yields lost output by industry and by month, including losses in industries whose own workers are unaffected but whose suppliers are.
Two known biases shape how we interpret the results. PEDSnet draws on large academic children's hospitals, which see sicker children and have better access to specialists. The adult RECOVER analysis found 38 significant post-acute diagnostic categories in New York City and only 11 in Florida, even though Florida is more deprived. The most plausible reading is that access to care drives how much gets documented. We therefore treat encounter-derived demand as a record of care consumed in well-served populations, not as a measure of need.
We bracket the estimate from both sides. The survey-based rate of persistent symptoms in the Beijing pediatric Omicron cohort serves as an upper bound: 55.0% of 2,001 children at 4–12 weeks, with concentration difficulty in 22.1%. That study had no uninfected comparison group, so we use it only as an upper scenario and never as a point estimate.
Baselines
The do-nothing reference is the original coupled model of Epidemiological and economic impact of COVID-19 in the US, run with its acute-only labor shock. Under its best mitigation scenario it reports an economic loss of about $3.4 trillion, and it ranks Professional Services among the hardest-hit industries. We rerun it under the same epidemic scenarios so that any additional loss and any change in ranking are attributable to the caregiver channel alone.
The strongest simpler competitor is a uniform caregiver multiplier. It adds post-acute parental absence as extra acute sick days, spreads them across industries in proportion to employment, and discounts them only by teleworkability. It uses the same PEDSnet demand totals and the same input-output model, so it isolates whether the timing and type structure of demand matters. No reported result for any pediatric caregiver labor channel was found on Lacuna, so there is no published figure for this baseline. We will report its output alongside ours.
The third baseline is a cost-of-illness calculation. It multiplies excess parent hours by industry wages, with no spread through supply chains, and so asks whether the input-output coupling adds anything. It parallels how the Brazil disability-adjusted life-years study measures burden: it totals years lived with disability without routing them through the economy.
Two ablations isolate the key ingredients. The first merges the three demand types into one, which tests whether time-fixed demand drives any ranking change. The second replaces the demand shapes of the individual diagnostic clusters with a single pooled shape. The proposal beats the uniform multiplier if two conditions hold together. The industry vulnerability ranking must differ meaningfully, with Kendall's rank correlation below 0.7 under the full model versus the multiplier. And that difference must persist across the entire sweep of $a_{jk}$ for time-fixed demand. It beats cost-of-illness if losses that spread through supply chains are a material share of total loss, which we set at more than 20% before running the analysis.
Why not just…?
Why not just scale acute absenteeism in the existing model? This is the real threat, and the uniform-multiplier baseline is built to test it directly. If time-fixed care demand is a small share of the total, or if schedule rigidity tracks teleworkability closely across industries, the multiplier will reproduce our ranking. In that case the correct conclusion is that existing models need only a longer absence parameter. We would report that outcome as the result rather than argue around it.
Why not just compute a cost-of-illness figure from parent hours and wages? Cost-of-illness gives a defensible total but cannot show that an industry with few affected parents still loses output because its suppliers do. The input-output model exists to capture exactly that spread. Our third baseline measures how large the spread is. If it is small, cost-of-illness is sufficient and we will say so.
Why not just use the survey prevalence of pediatric symptoms instead of health records? Survey prevalence records whether a child has symptoms, not how much parental time those symptoms demand or when. It also lacks controls. The Beijing cohort's 55% has no uninfected comparison group, and fever alone was reported by 95.1% of infected children, which suggests how broadly symptom surveys cast their net. Encounter streams measured against matched controls give demand shaped in time that can be attributed to infection. The survey figure is kept as an upper-bound scenario.
Why not just forecast industry employment with a time-series model? MacroBench shows that a simple autoregressive model, which predicts each month from the previous one, is hard to beat on aggregate series. Only one frontier language model narrowly outperformed it once leakage of historical dates into the models was audited. But a forecast of total employment cannot attribute a shortfall to a mechanism or rank industries by exposure to a specific channel. That attribution and ranking are what this proposal delivers.
Proposed Experiments
The first check is whether caregiver demand has structure worth modeling. In PEDSnet, we test whether excess encounters over months 1–12 differ by demand type and by diagnostic cluster between children with long COVID and infected controls. We require confidence intervals on the time-fixed share that exclude a pooled common value. If all clusters share the same demand shape, the second ablation will show no effect and the model simplifies.
The second check is the ranking-reversal test. We run the full model, the uniform multiplier, and cost-of-illness on the same epidemic scenarios. For each, we report output lost by industry, the Kendall correlation between rankings, and the share of loss that spreads through supply chains, all across the full sweep of $a_{jk}$ and across three pediatric burden scenarios. The lower bound comes from infection-attributable demand in PEDSnet, the middle from total excess demand in PEDSnet, and the upper from demand scaled to the Beijing prevalence.
The decisive result is the direction of change in healthcare's position. If healthcare moves toward the top of the vulnerability ranking under the full model but not under the multiplier, the mechanism we propose is carrying the result.
The third check is external. We compare predicted industry-specific parental absence against observed industry absence rates from a labor-force survey during pediatric infection waves. That series is outside the research file, and no reported result was found on Lacuna for this comparison. We treat the match as a test the model could fail. We will report the correlation between predicted and observed absence across industries, and we commit to reporting the model as unsupported if its industry ranking of absence does no better than the uniform multiplier's.
References
- Epidemiological and economic impact of COVID-19 in the US
- COVID-EENet: Predicting Fine-Grained Impact of COVID-19 on Local Economies
- Understanding pediatric long COVID using a tree-based scan statistic approach: An EHR-based cohort study from the RECOVER Program
- Characterization of long COVID temporal sub-phenotypes by distributed representation learning from electronic health record data: a cohort study
- Impact of SARS-CoV-2 Infection on Disease Trajectory in Youth with T1D: An EHR-Based Cohort Study from the RECOVER Program
- Data-driven analysis to understand long COVID using electronic health records from the RECOVER initiative
- 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
- Disability-adjusted life years associated with COVID-19 in Brazil, 2020
- MacroBench: Measuring Frontier LLM Macroeconomic Forecasting Ability