Auto-Research PipelineA durable, human-gated pipeline for computational research

← All examples

Research run10fc99e05068· 24 October 2026

Do children under 5 years have higher SARS-CoV-2 upper-respiratory viral loads (lower RT-PCR cycle threshold values) than adults, once testing setting and calendar time are accounted for?

CompleteArchived runEpidemiology & Public Health · Modelled on American Journal of Epidemiology

Paper

Infant–Adult Differences in SARS-CoV-2 Cycle Threshold Values and the Limits of Deidentified Testing Records: A Cross-Sectional Analysis

Abstract

Young children are often said to carry more severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) than adults, but comparisons rarely consider why and when people were tested. In a cross-sectional analysis of 15,524 deidentified reverse-transcription polymerase chain reaction tests from the Children's Hospital of Philadelphia (2020), we studied each person's first positive specimen with a quantifiable cycle threshold, or Ct (785 persons). Lower Ct means more RNA. Within age bands, recorded age was ordered by date of first test (Spearman ρ = 0.975 at ages 1–17 years) and never changed within 1,744 repeat testers. Only 1 of 33 clinics had its own weekday pattern, about what chance would give, and recorded dates showed no association with Ct, so neither ages 1–4 years nor calendar time could be analyzed. Without recorded dates, infants' Ct was 1.72 cycles lower than adults' unadjusted (95% confidence interval: −3.65, 0.20) and 2.11 cycles lower with setting adjustment (95% confidence interval: −4.04, −0.18), about 4.3-fold more RNA. Infant-specific selection on viral load (odds ratio 1.20 per 5 cycles) would remove it. These records are compatible with a small infant excess but cannot show whether children under 5 years differ from adults once calendar time is considered.

KeywordsSARS-CoV-2cycle thresholdviral loadinfantsselection biasquantitative bias analysisdeidentificationcross-sectional studies

§Ask about this paper

Grounded in this run

Questions are answered from this run's record — the paper, the executed analysis, the data provenance and the referee report. Nothing else is consulted.

This run is finished, so the recorded paper cannot be rewritten. Questions are still answered from its record.

Pipeline

1.Literature review

Attempt 3Done

Literature review complete (live:agent:claude-opus-5-5): 10 key papers, 46 grounded sources, 5 open gaps.

Summary

The view that children under 5 carry more SARS-CoV-2 in the upper respiratory tract than adults rests mainly on a small study. Heald-Sargent and colleagues studied 145 symptomatic patients at a Chicago children's hospital and found median Ct values of 6.5 at <5 years versus 11.0 in adults, which they read as a 10- to 100-fold RNA excess. Larger laboratory and population studies disagree. Euser and colleagues analysed first specimens of 25,374 positive patients in Int J Epidemiol and found viral load rising with age, with children under 12 lower than adults. Jones and colleagues analysed 25,381 German cases and found the youngest subjects lower by 0.5 log10 or less. Madera and colleagues and Baggio and colleagues found no difference, and Bullard and colleagues found higher Ct and less culturable virus in children. In Seattle community surveillance, Chung and colleagues found symptom status shifted Ct by about 3 cycles in both children and adults, while symptomatic children and symptomatic adults did not differ (-0.7; 95% CI -2.2 to 0.9). Two lines of work explain much of this disagreement. First, testing indication matters. In nine paediatric hospital programmes, Kociolek and colleagues found that asymptomatic children's adjusted Ct was 10.3 cycles higher than symptomatic children's, and at the Children's Hospital of Philadelphia, Otto and colleagues documented testing criteria that moved from symptomatic testing to preprocedural and universal preadmission screening within weeks, across drive-through, emergency and inpatient sites. Second, calendar time matters. Hay and colleagues and Walker and colleagues showed that Ct distributions among positives shift with epidemic phase, so age groups tested at different times are not comparable without adjustment for date. Infant-specific studies point to heterogeneity inside 'under 5'. Ochoa and colleagues found the highest loads in infants younger than 6 months, with older children similar to adults, and Pinninti and colleagues found much higher respiratory loads in hospitalised infants <1 year. Ct is also assay- and specimen-specific (Rhoads and colleagues), and culture positivity is rare above Ct 24 (Bullard and colleagues, 2020). Methodologically, Ct is observable only in people who were tested and tested positive. Selection on symptoms, and hence on viral load, therefore creates the collider structure described by Hernan, Hernandez-Diaz and Robins and by Griffith and colleagues for COVID-19. Standard tools exist to quantify such biases: quantitative bias analysis (Greenland; Lash and colleagues), E-values (VanderWeele and Ding), omitted-variable-bias robustness values for regression coefficients (Cinelli and Hazlett), and exponential-tilt selection models (Scharfstein, Rotnitzky and Robins). They are rarely applied to Ct contrasts. Prevalence ratios for binary viral-load markers are estimated by modified Poisson regression (Zou), and STROBE and its RECORD extension govern reporting of routinely collected laboratory data. Finally, the open CHOP teaching dataset used here is documented only as 'anonymized, time-shifted, and permuted'. Disclosure-control methods such as data swapping (Dalenius and Reiss) can distort joint distributions in hard-to-predict ways, and the epidemiologic literature on checking whether associations survive such processing is thin.

Key papers (10)

  1. Age-Related Differences in Nasopharyngeal Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) Levels in Patients With Mild to Moderate Coronavirus Disease 2019 (COVID-19). Heald-Sargent T, Muller WJ, Zheng X, Rippe J, Patel AB, Kociolek LK; JAMA Pediatrics 174(9):902-903, 2020.
    The source of the under-5 hypothesis: 145 symptomatic patients tested within 1 week of onset at a Chicago children's hospital. Median Ct was 6.5 at <5 y, 11.1 at 5-17 y and 11.0 at 18-65 y (about 10- to 100-fold more RNA). No adjustment for setting or date.
  2. SARS-CoV-2 viral-load distribution reveals that viral loads increase with age: a retrospective cross-sectional cohort study. Euser S, Aronson S, Manders I, et al.; International Journal of Epidemiology 50(6):1795-1803, 2022.
    The closest design and venue: first specimen per patient in routinely collected laboratory records (25,374 positives). Viral load increased with age, with children <12 y lower than adults (beta -0.03; 95% CI -0.03 to -0.02). Median Cp differed by more than 4 cycles between >79 and <12 y.
  3. Estimating infectiousness throughout SARS-CoV-2 infection course. Jones TC, Biele G, Muhlemann B, et al.; Science 373(6551):eabi5273, 2021.
    25,381 German cases across test centres and hospitals. The youngest subjects' viral load was lower by 0.5 log10 or less, partly because of smaller swabs. A large-sample counterpoint to the under-5 claim.
  4. Comparison of Symptoms and RNA Levels in Children and Adults With SARS-CoV-2 Infection in the Community Setting. Chung E, Chow EJ, Wilcox NC, et al.; JAMA Pediatrics 175(10):e212025, 2021.
    Community cross-sectional study of 555 positives. Symptomatic people had Ct about 3 cycles lower than asymptomatic people. Symptomatic children vs symptomatic adults: adjusted -0.7 (95% CI -2.2 to 0.9). Shows that testing indication confounds age contrasts.
  5. Infants Younger Than 6 Months Infected With SARS-CoV-2 Show the Highest Respiratory Viral Loads. Ochoa V, Diaz FE, Ramirez E, et al.; Journal of Infectious Diseases 225(3):392-395, 2022.
    45,318 positive NP swabs: infants <6 months had higher viral loads than any other age group, and children older than 6 months were similar to adults. This places any under-5 excess in infancy.
  6. Virological Characteristics of Hospitalized Children With SARS-CoV-2 Infection. Pinninti SG, Pati S, Poole C, et al.; Pediatrics 147(5):e2020037812, 2021.
    102 hospitalised children screened at admission: respiratory viral load was much higher at <1 y than in older children (3.3x10^7 vs 1.3x10^4 copies/mL).
  7. Comparison of Upper Respiratory Viral Load Distributions in Asymptomatic and Symptomatic Children Diagnosed with SARS-CoV-2 Infection in Pediatric Hospital Testing Programs. Kociolek LK, Muller WJ, Yee R, et al.; Journal of Clinical Microbiology 59(1):e02593-20, 2020.
    Nine paediatric hospital programmes: asymptomatic children's adjusted Ct was 10.3 cycles higher than symptomatic children's in every age bracket, and surveillance testing was associated with lowest-quartile Ct. Testing setting and indication are first-order confounders.
  8. The Epidemiology of Severe Acute Respiratory Syndrome Coronavirus 2 in a Pediatric Healthcare Network in the United States. Otto WR, Geoghegan S, Posch LC, et al.; Journal of the Pediatric Infectious Diseases Society 9(5):523-529, 2020.
    Describes the CHOP testing programme that generated the study data: drive-through, emergency, outpatient, inpatient and urgent-care testing. Criteria moved from symptomatic testing to preprocedural (24 March) and universal preadmission (1 April 2020) screening. Positivity was 21.1% with symptoms or exposure vs 3.8% at screening. Assays were a laboratory-developed N2 test and Xpert.
  9. Estimating epidemiologic dynamics from cross-sectional viral load distributions. Hay JA, Kennedy-Shaffer L, Kanjilal S, et al.; Science 373(6552):eabh0635, 2021.
    Shows that the Ct distribution among positives shifts with epidemic phase. Calendar time therefore confounds any comparison of groups tested at different times.
  10. Predicting Infectious Severe Acute Respiratory Syndrome Coronavirus 2 From Diagnostic Samples. Bullard J, Dust K, Funk D, et al.; Clinical Infectious Diseases 71(10):2663-2666, 2020.
    No culture growth from samples with Ct >24 or symptom-to-test >8 days (OR 0.64 per Ct unit). The basis for a Ct <=24 marker of likely culturable virus.

Methods in use

  • Ct analysed as a continuous outcome: medians/IQR with rank tests (Heald-Sargent; Bullard 2021), adjusted linear regression of Ct (Chung; Euser), quantile regression of median Ct (Walker), kinetic models (Jones; Bellon).
  • Ct differences translated into fold differences in RNA by assuming one doubling per cycle (Heald-Sargent: 10- to 100-fold; Euser: 4 cycles is about 16-fold), with assay- and institution-specific centring when pooling platforms (Kociolek).
  • Binary viral-load markers anchored to culture (Ct <=24, Bullard 2020), with prevalence ratios from modified Poisson regression with robust variance (Zou, Am J Epidemiol 2004) and heteroskedasticity-consistent HC3 standard errors (MacKinnon and White).
  • Confounding control by symptom status or duration, sex, testing indication and restriction to a symptom-onset window (Heald-Sargent; Baggio); joint adjustment for setting and calendar date is rare. Covariate coefficients are not interpreted as effects (Westreich and Greenland's Table 2 fallacy).
  • Structural (collider) view of selection into testing and positivity (Hernan, Hernandez-Diaz and Robins 2004; Griffith et al. 2020; Accorsi et al. 2021), with bounding (Smith and VanderWeele 2019) and exponential-tilt selection models (Scharfstein, Rotnitzky and Robins 1999).
  • Quantitative bias analysis for unmeasured confounding: bias formulas (Greenland 1996; Lash et al. 2014), E-values (VanderWeele and Ding 2017), and omitted-variable-bias robustness values with partial-R2 benchmarking for linear-regression coefficients (Cinelli and Hazlett 2020).
  • Reporting of routinely collected laboratory data under STROBE (von Elm et al. 2007) and its RECORD extension (Benchimol et al. 2015).

Open gaps

  • Few Ct-by-age studies adjust jointly for testing setting and calendar time, even though both are known to shift Ct distributions (Kociolek; Otto; Hay; Walker). The adjusted under-5 versus adult contrast in hospital testing records is therefore not settled.
  • 'Under 5' pools infants and toddlers. Infant-specific evidence (Ochoa; Pinninti) suggests the excess, if any, lies in infancy, but infant versus adult contrasts within one laboratory, adjusted for setting and date, are rare.
  • Selection into testing and positivity depends on symptoms and hence on viral load, differently by age. This collider structure is named but almost never quantified for Ct contrasts, for example with an exponential-tilt or inverse-probability-of-selection bias analysis.
  • Open deidentified teaching datasets are released 'anonymized, time-shifted, and permuted' without saying which fields were altered. Whether the age-outcome structure survives such processing is rarely checked before measures of association are estimated, and disclosure-control work (Dalenius and Reiss; Kim) warns that it may not.
  • Adults tested at a paediatric hospital (staff, family members, outside referrals, adult patients) are an unusual comparator whose setting mix differs sharply from the children's. Within-setting comparisons, for example drive-through only, and restriction to the hospital's own patients are needed to tell an age difference from a difference in why people were tested.
Grounded search sources (46) · mode: live:agent:claude-opus-5-5

2.Ideation

Attempt 2Done

3 candidate ideas generated; selected: “Do children under 5 have lower SARS-CoV-2 cycle thresholds than adults once testing setting and calendar time are accounted for? An identifiability-first cross-sectional analysis of deidentified paediatric-hospital testing records”.

  1. Idea 1Selected

    Do children under 5 have lower SARS-CoV-2 cycle thresholds than adults once testing setting and calendar time are accounted for? An identifiability-first cross-sectional analysis of deidentified paediatric-hospital testing records

    Hypothesis. Among people testing positive by RT-PCR at the Children's Hospital of Philadelphia in 2020, children younger than 5 years have a lower mean cycle threshold (Ct) than adults aged 18 years or older, i.e. more upper-respiratory viral RNA, after adjustment for testing setting and calendar time. The infant-specific literature (Ochoa 2022; Pinninti 2021) predicts the excess will be clearest in infants younger than 1 year. Because the records were 'anonymized, time-shifted, and permuted', the hypothesis is tested only on under-5 age bands whose membership is shown to survive deidentification.

    Dataset. COVID-19 PCR testing at the Children's Hospital of Philadelphia, 2020 chop_covid_testing

    Test plan, novelty and risk

    Test plan. Design: cross-sectional analysis of routinely collected laboratory records, reported to STROBE/RECORD. Sample: start from all 15,524 tests; drop invalid results; keep positive results with a quantifiable Ct (non-missing and below 45, the not-detected placeholder); drop recorded ages above 100 years; keep each person's first quantifiable positive specimen (sorted by collection day). Exposure: recorded age band, infant <1 y, 1-4 y, 5-11 y, 12-17 y, and adult >=18 y as reference; composite under-5 = infant + 1-4 y. Step 1, identifiability (pre-specified): (a) Spearman correlation between recorded age and each person's first test day within the infant, 1-17 y and adult bands, over all tests; (b) checks against known clinical structure: share of NICU tests recorded <1 y, share of employee-health tests recorded >=18 y, share of OB-GYN tests recorded female; (c) whether Ct keeps its association with setting and calendar time, judged by R-squared against a 1,000-permutation null. Decision rule: where recorded age is near-monotone in first test day inside a band (rho >= 0.9), sub-bands of that band are treated as calendar-time markers. The under-5 hypothesis is then tested through the preserved bands (infants vs adults, children 1-17 y vs adults), and the composite under-5 and 1-4 y contrasts are reported descriptively. Step 2, measures of association: mean Ct difference versus adults by OLS with HC3 95% confidence intervals in three nested models. M1 is unadjusted. M2 adds calendar time as a restricted cubic spline of collection day with 4 knots at the 5th, 35th, 65th and 95th percentiles. M3 adds testing setting, derived from drive_thru_ind, demo_group, patient_class and clinic_name: drive-through (reference), emergency department/urgent care, inpatient including NICU/PICU and observation, outpatient/other, and outside referring institution. M3 is fitted for the preserved-band coding (infant, child 1-17 y), the composite coding (under 5, 5-17 y) and the recorded five-band coding. The infant-vs-adult difference is translated to a fold difference in RNA as 2^(-difference), assuming one doubling per cycle. Secondary outcome: Ct <= 24, a marker of likely culturable virus, as a prevalence ratio from modified Poisson regression with robust variance (M1 and M3), with an E-value. Effect-measure modification: infant-vs-adult difference within the drive-through and within the other settings, plus an infant x drive-through interaction. Sensitivity analyses: weekly indicators instead of the spline; all quantifiable positive specimens with person-clustered SEs; excluding outside institutions; the hospital's own patients only (demo_group 'patient'). Quantitative bias analysis: (i) unmeasured confounding, for example by days since symptom onset: Cinelli-Hazlett robustness values for the M3 infant coefficient, set against the partial R-squared of setting with Ct and with infant status; (ii) selection on viral load: inverse-probability-of-selection weights under an exponential-tilt selection model in which the probability of being tested and detected rises by an odds ratio of 1.25, 1.5 or 2 per 5-cycle lower Ct, applied to infants or to adults, plus the tipping-point infant odds ratio at which the adjusted difference reaches zero. Refutation: the primary hypothesis (infants vs adults, M3) is refuted if the upper 95% confidence limit of the adjusted difference is >= 0; the composite under-5 hypothesis is refuted if its M3 upper limit is >= 0; the 1-4 y contrast is declared not identifiable if the identifiability check fails. hypothesis_supported is defined as an M3 infant-vs-adult upper limit below 0.

    Novelty. The literature disagrees: Heald-Sargent and Strutner find lower Ct under 5, Madera, Baggio and Chung find no difference, and Euser, Jones and Bullard find higher Ct in children. Few studies adjust jointly for testing setting and calendar time, although Kociolek, Otto, Hay and Walker show that both move Ct distributions. Infant versus adult contrasts within a single laboratory are rare, and selection on viral load into testing is named but not quantified. No analysis of this widely used teaching dataset has asked whether its age structure survives the documented permutation before estimating an age association. The study combines all four: an identifiability check, joint setting and time adjustment, an infant-specific contrast, and quantitative bias analysis for both unmeasured confounding and selection.

    Expected outcome. From the infant literature: infants have a lower adjusted mean Ct than adults by a few cycles, a several-fold RNA excess, with a wide interval because infants are a small group, while children 1-17 y as a whole differ little from adults. Whether the 1-4 y band can be assessed depends on the identifiability check. If recorded ages within 1-17 y track calendar time, the composite under-5 estimate will mix infants with a calendar-period marker and will not answer the question for toddlers. The bias analyses are expected to show that an adjusted difference of a few cycles is sensitive to plausible differential selection on viral load.

    Main risk. Undocumented deidentification may also have altered Ct, the setting fields or band membership, and only its consequences, not its procedure, can be checked. There is no information on symptoms, days since onset, specimen type or assay platform. Selection into testing and positivity plausibly depends on viral load differently by age and setting, a collider structure. Fewer than 100 infants will be analysed. Adults tested at a paediatric hospital (staff, family members, outside referrals) are an unusual comparison group.

  2. Idea 2

    Effect-measure modification of the child-adult Ct difference by testing setting: screening-dominated versus symptom-driven sites

    Hypothesis. The difference in mean Ct between children (<18 y) and adults among first quantifiable positive specimens is larger, with children lower relative to adults, in symptom-driven settings (emergency department, drive-through) than in screening-dominated settings (inpatient preadmission and preprocedural testing), after adjustment for calendar time.

    Dataset. COVID-19 PCR testing at the Children's Hospital of Philadelphia, 2020 chop_covid_testing

    Test plan, novelty and risk

    Test plan. Same sample construction as the primary study: positive results with Ct <45, first quantifiable specimen per person, ages <=100 y. Settings are derived from drive_thru_ind, patient_class, clinic_name and demo_group. OLS of Ct on child status, setting and their interaction, adjusted for a restricted cubic spline of collection day, with HC3 95% CIs. Report setting-specific child-adult differences and the interaction contrast (symptom-driven minus screening-dominated). Refuted if the interaction's 95% CI includes 0 or the contrast has the opposite sign.

    Novelty. Kociolek showed that asymptomatic children's Ct is about 10 cycles higher than symptomatic children's, and Chung showed that symptom status rather than age moves Ct. No study has estimated whether the age contrast itself is modified by testing setting within one laboratory.

    Expected outcome. Higher Ct in screening-dominated inpatient settings for all ages. Any child-adult difference is expected to be imprecise outside the drive-through because few adults are tested in emergency and inpatient settings at a paediatric hospital.

    Main risk. Few adults outside the drive-through give very wide setting-specific intervals. Setting is itself affected by age and illness severity, a potential collider. Indication (screening vs symptomatic) is not recorded and is only proxied by setting.

  3. Idea 3

    Calendar-time confounding of Ct contrasts: does the Ct distribution among positives track the epidemic trajectory in a paediatric hospital?

    Hypothesis. Weekly median Ct among first quantifiable positive specimens is inversely associated with the weekly growth rate of test positivity: lower Ct when positivity is rising, higher when falling, as predicted by Hay et al. (2021). Calendar time is therefore a confounder of between-group Ct contrasts.

    Dataset. COVID-19 PCR testing at the Children's Hospital of Philadelphia, 2020 chop_covid_testing

    Test plan, novelty and risk

    Test plan. Aggregate valid tests (excluding invalid results) by collection week. Compute weekly positivity, the growth rate of log positivity between consecutive weeks, and the weekly median Ct among first quantifiable positive specimens. Regress weekly median Ct on growth rate by OLS with HC3 standard errors, weighting by the number of positives per week. Refuted if the slope's 95% CI includes 0 or is positive.

    Novelty. Hay et al. and Walker et al. showed epidemic-phase effects on Ct in population surveillance. Whether the same signal appears in a single paediatric hospital, where testing policy changed quickly, is untested, and it determines whether date adjustment is needed when comparing age groups.

    Expected outcome. A negative but imprecise slope.

    Main risk. Only about 15 weekly observations. Changes in testing policy (screening expansion) move positivity independently of epidemic growth. Time-shifting of the dates is undocumented.

Ranking rationale

Idea 0 answers the stated question directly, with design features the literature lacks: an identifiability check of the deidentified age field, joint adjustment for setting and calendar time, an infant-specific contrast, and quantitative bias analysis for confounding and selection. It uses the full individual-level sample. Idea 1 is a narrower effect-modification question that idea 0 partly contains and that the small number of adults outside the drive-through would leave imprecise. Idea 2 rests on about 15 weekly observations and answers a methodological side question about calendar-time confounding rather than the age question.

3.Experimentation

Attempt 6Done

Experiment succeeded in 1 round (1.465s).

Reviewer feedback applied: Independent referee (Claude Fable 5.1), round 2: minor_revision. Required changes: (1) [analysis] Re-run the within-setting contrasts, the infant x drive-through product term, the three applicable sensitivity analyses (all quantifiable specimens with person-clustered SEs, excluding outside institutions, hospital patients only) and the selection-OR scenarios under the setting-only specification, and compute setting's partial R-squared with Ct and with infant status without the spline as the benchmark for the setting-only robustness values; report these in Table 4 in place of, or beside, the M3 versions, and label the weekly-indicator row as a check that the time coding is immaterial. (2) [text] In Results and the Abstract, state the weekday evidence as 1 of 33 clinics departing from the overall profile (about what chance would give), drop or explicitly discount the emergency-department vs drive-through comparison since both operate daily, and add that the Ct-date check has limited power (3-df spline, 785 specimens) so the detachment conclusion rests mainly on the uniform profiles of weekday-only clinics. (3) [text] Qualify the Yonker citation in the Introduction: no age gradient within that pediatric cohort, but early viral loads above those of hospitalized adults, so cite it for the former rather than as a no-difference study.
Succeeded1 round · exit 0 · 1.465s

Data provenance

Datasetchop_covid_testing

COVID-19 PCR testing at the Children's Hospital of Philadelphia, 2020

Source
R package 'medicaldata', mirrored by the Rdatasets archive; deidentified records from the Children's Hospital of Philadelphia
Citation
Higgins, P.D.R. (2021) medicaldata: Data Package for Medical Datasets. R package version 0.2.0. https://CRAN.R-project.org/package=medicaldata. Data donated by Amrom E. Obstfeld; deidentified, time-shifted and value-permuted records of SARS-CoV-2 RT-PCR testing at the Children's Hospital of Philadelphia in 2020. There is no source publication for this dataset -- the package's source field is a data-donation acknowledgement only.
Licence
Redistributed in the R package 'medicaldata' under MIT (LICENSE file: copyright 2021, medicaldata authors) via the Rdatasets archive. No separate data licence is stated. The records are real patient data that have been deidentified, time-shifted and permuted -- not synthetic.
File
https://vincentarelbundock.github.io/Rdatasets/csv/medicaldata/covid_testing.csv
Observations
15,524
Variables (15)
subject_id, gender, pan_day, test_id, clinic_name, result, demo_group, age, drive_thru_ind, ct_result, orderset, payor_group, patient_class, col_rec_tat, rec_ver_tat
Units
pan_day is the day of the pandemic period on which the test was collected, running from 4 to 107. result is 'negative' (14,358), 'positive' (865) or 'invalid' (301). ct_result is the RT-PCR cycle threshold, where lower values mean higher viral load and the value is only interpretable for positive tests. age is in years (0 for infants). col_rec_tat is collection-to-receipt turnaround and rec_ver_tat receipt-to-verification turnaround, both in hours. drive_thru_ind and orderset are 0/1 indicators.
Retrieved
24 Oct 2026, 03:07
Checksum
sha256 fc7d7c522749144b…

Real, publicly available data — downloaded and checksum-verified before the analysis ran, never simulated.

Result (RESULT_JSON)

{
  "design": "cross-sectional; first quantifiable positive specimen per person",
  "flow": {
    "tests": 15524,
    "persons": 12344,
    "invalid": 301,
    "positive_tests": 865,
    "quantifiable_positive_tests": 857,
    "after_age_le_100": 856,
    "analytic_persons": 785,
    "persons_with_2plus_quantifiable_positives": 45
  },
  "deidentification": {
    "spearman_age_vs_first_test_day": {
      "adult": 0.644,
      "child": 0.975,
      "infant": 0.801
    },
    "median_recorded_age_by_first_test_period": {
      "adult": {
        "days4_37": 25,
        "days38_72": 33,
        "days73_107": 47
      },
      "child": {
        "days4_37": 2,
        "days38_72": 6,
        "days73_107": 14
      },
      "infant": {
        "days4_37": 0.1,
        "days38_72": 0.4,
        "days73_107": 0.8
      }
    },
    "persons_by_band": {
      "child": 7179,
      "adult": 3665,
      "infant": 1500
    },
    "r2_recorded_age_on_time_children_1_17": 0.671,
    "clinical_structure_checks": {
      "nicu_tests": 294,
      "nicu_pct_age_under1": 97.6,
      "employee_health_tests": 12,
      "employee_health_pct_age18plus": 100,
      "obgyn_tests": 69,
      "obgyn_pct_female": 56.5
    }
  },
  "ct_integrity": {
    "r2_setting_time": 0.055,
    "perm_null_p95": 0.018,
    "perm_null_p99": 0.022,
    "perm_share_ge_observed": 0
  },
  "spline_knots_day": [
    19,
    47,
    78,
    103.8
  ],
  "table1": {
    "infant": {
      "n": 79,
      "ct_mean": 26.94,
      "ct_sd": 7.96,
      "ct_median": 24.32,
      "ct_q1": 20.6,
      "ct_q3": 34.68,
      "pct_ct_le24": 46.8,
      "day_median": 52,
      "day_q1": 30,
      "day_q3": 71,
      "pct_drive": 27.8,
      "pct_emergency": 38,
      "pct_inpatient": 6.3,
      "pct_outpatient": 15.2,
      "pct_outside": 12.7
    },
    "1-4": {
      "n": 107,
      "ct_mean": 30.85,
      "ct_sd": 7.74,
      "ct_median": 33.56,
      "ct_q1": 23.31,
      "ct_q3": 37.8,
      "pct_ct_le24": 28,
      "day_median": 29,
      "day_q1": 23,
      "day_q3": 35,
      "pct_drive": 42.1,
      "pct_emergency": 29,
      "pct_inpatient": 15.9,
      "pct_outpatient": 13.1,
      "pct_outside": 0
    },
    "5-11": {
      "n": 117,
      "ct_mean": 28.92,
      "ct_sd": 7.6,
      "ct_median": 28.19,
      "ct_q1": 22.18,
      "ct_q3": 36.2,
      "pct_ct_le24": 34.2,
      "day_median": 53,
      "day_q1": 44,
      "day_q3": 63,
      "pct_drive": 49.6,
      "pct_emergency": 33.3,
      "pct_inpatient": 2.6,
      "pct_outpatient": 12.8,
      "pct_outside": 1.7
    },
    "12-17": {
      "n": 168,
      "ct_mean": 29.36,
      "ct_sd": 6.91,
      "ct_median": 28.57,
      "ct_q1": 23.51,
      "ct_q3": 36.09,
      "pct_ct_le24": 27.4,
      "day_median": 84.5,
      "day_q1": 72.8,
      "day_q3": 97,
      "pct_drive": 45.2,
      "pct_emergency": 26.2,
      "pct_inpatient": 2.4,
      "pct_outpatient": 6,
      "pct_outside": 20.2
    },
    "adult": {
      "n": 314,
      "ct_mean": 28.66,
      "ct_sd": 6.91,
      "ct_median": 27.93,
      "ct_q1": 22.74,
      "ct_q3": 35.02,
      "pct_ct_le24": 31.2,
      "day_median": 70.5,
      "day_q1": 43.2,
      "day_q3": 91,
      "pct_drive": 79.9,
      "pct_emergency": 5.7,
      "pct_inpatient": 2.2,
      "pct_outpatient": 3.2,
      "pct_outside": 8.9
    },
    "all": {
      "n": 785,
      "ct_mean": 28.97,
      "ct_sd": 7.29,
      "ct_median": 28.36,
      "ct_q1": 22.6,
      "ct_q3": 35.87,
      "pct_ct_le24": 32,
      "day_median": 63,
      "day_q1": 38,
      "day_q3": 86,
      "pct_drive": 57.6,
      "pct_emergency": 20.6,
      "pct_inpatient": 4.6,
      "pct_outpatient": 7.8,
      "pct_outside": 9.4
    }
  },
  "association_mean_ct_diff_vs_adults": {
    "preserved_M1": {
      "infant": {
        "diff": -1.72,
        "lo": -3.65,
        "hi": 0.2
      },
      "child": {
        "diff": 0.97,
        "lo": -0.09,
        "hi": 2.03
      }
    },
    "preserved_M2": {
      "infant": {
        "diff": -1.94,
        "lo": -3.88,
        "hi": 0
      },
      "child": {
        "diff": 0.93,
        "lo": -0.13,
        "hi": 1.99
      }
    },
    "preserved_M3": {
      "infant": {
        "diff": -2.25,
        "lo": -4.2,
        "hi": -0.31
      },
      "child": {
        "diff": 0.72,
        "lo": -0.4,
        "hi": 1.85
      }
    },
    "composite_M1": {
      "u5": {
        "diff": 0.53,
        "lo": -0.86,
        "hi": 1.92
      },
      "a5_17": {
        "diff": 0.52,
        "lo": -0.62,
        "hi": 1.65
      }
    },
    "composite_M2": {
      "u5": {
        "diff": -0.13,
        "lo": -1.6,
        "hi": 1.34
      },
      "a5_17": {
        "diff": 0.8,
        "lo": -0.35,
        "hi": 1.95
      }
    },
    "composite_M3": {
      "u5": {
        "diff": -0.52,
        "lo": -2.06,
        "hi": 1.03
      },
      "a5_17": {
        "diff": 0.74,
        "lo": -0.46,
        "hi": 1.94
      }
    },
    "recorded_M1": {
      "infant": {
        "diff": -1.72,
        "lo": -3.65,
        "hi": 0.2
      },
      "a1_4": {
        "diff": 2.19,
        "lo": 0.53,
        "hi": 3.85
      },
      "a5_11": {
        "diff": 0.26,
        "lo": -1.32,
        "hi": 1.84
      },
      "a12_17": {
        "diff": 0.7,
        "lo": -0.6,
        "hi": 1.99
      }
    },
    "recorded_M2": {
      "infant": {
        "diff": -1.92,
        "lo": -3.87,
        "hi": 0.02
      },
      "a1_4": {
        "diff": 1.53,
        "lo": -0.31,
        "hi": 3.36
      },
      "a5_11": {
        "diff": 0.36,
        "lo": -1.34,
        "hi": 2.05
      },
      "a12_17": {
        "diff": 0.94,
        "lo": -0.41,
        "hi": 2.29
      }
    },
    "recorded_M3": {
      "infant": {
        "diff": -2.2,
        "lo": -4.16,
        "hi": -0.24
      },
      "a1_4": {
        "diff": 1.13,
        "lo": -0.78,
        "hi": 3.05
      },
      "a5_11": {
        "diff": 0.75,
        "lo": -0.99,
        "hi": 2.49
      },
      "a12_17": {
        "diff": 0.51,
        "lo": -0.89,
        "hi": 1.91
      }
    }
  },
  "primary_infant_vs_adult_M3": {
    "diff": -2.25,
    "lo": -4.2,
    "hi": -0.31
  },
  "fold_rna_infant_vs_adult": {
    "fold": 4.76,
    "lo": 1.24,
    "hi": 18.38
  },
  "prevalence_ratio_ct_le24": {
    "infant_M1": {
      "pr": 1.5,
      "lo": 1.13,
      "hi": 2
    },
    "infant_M3": {
      "pr": 1.51,
      "lo": 1.11,
      "hi": 2.06
    },
    "evalue_point": 2.39,
    "evalue_ci": 1.46,
    "u5_M3": {
      "pr": 1.27,
      "lo": 0.95,
      "hi": 1.7
    }
  },
  "effect_modification_setting": {
    "drive_through": {
      "diff": -2.73,
      "lo": -5.77,
      "hi": 0.3
    },
    "other_settings": {
      "diff": -1.9,
      "lo": -4.79,
      "hi": 0.99
    },
    "n_infant_drive": 22,
    "n_adult_drive": 251,
    "interaction_infant_x_drive": {
      "diff": -1.08,
      "lo": -5.22,
      "hi": 3.07
    }
  },
  "sensitivity_infant_vs_adult": {
    "weekly_time": {
      "diff": -2.25,
      "lo": -4.2,
      "hi": -0.3
    },
    "all_specimens_cluster": {
      "diff": -1.79,
      "lo": -3.63,
      "hi": 0.06
    },
    "no_outside_institutions": {
      "diff": -2.56,
      "lo": -4.66,
      "hi": -0.45
    },
    "chop_patients_only": {
      "diff": -2.83,
      "lo": -5.32,
      "hi": -0.33
    },
    "n_all_specimens": 856,
    "n_patients_only": 484,
    "n_adults_patients_only": 76
  },
  "bias_unmeasured_confounding": {
    "partial_r2_infant_ct": 0.0072,
    "rv_point": 0.0818,
    "rv_ci": 0.0147,
    "partial_r2_setting_ct": 0.0488,
    "partial_r2_setting_infant": 0.1022
  },
  "bias_selection_tilt": {
    "inf1.25_ad1": {
      "diff": 0.32,
      "lo": -1.72,
      "hi": 2.37
    },
    "inf1.5_ad1": {
      "diff": 2.33,
      "lo": 0.31,
      "hi": 4.34
    },
    "inf2_ad1": {
      "diff": 4.84,
      "lo": 3,
      "hi": 6.68
    },
    "inf1_ad1.5": {
      "diff": -5.74,
      "lo": -7.66,
      "hi": -3.83
    },
    "inf1_ad2": {
      "diff": -7.63,
      "lo": -9.52,
      "hi": -5.75
    },
    "tipping_or_infant": 1.22
  },
  "hypothesis_supported": true,
  "repeat_testers": {
    "persons": 1744,
    "tests": 4924,
    "pct_constant_age": 100,
    "by_band": {
      "adult": {
        "persons": 483,
        "pct_constant_age": 100,
        "span_median": 35,
        "rho_age_first_day": 0.551,
        "rho_age_last_day": 0.149
      },
      "child": {
        "persons": 938,
        "pct_constant_age": 100,
        "span_median": 39,
        "rho_age_first_day": 0.903,
        "rho_age_last_day": 0.276
      },
      "infant": {
        "persons": 323,
        "pct_constant_age": 100,
        "span_median": 43,
        "rho_age_first_day": 0.431,
        "rho_age_last_day": 0.458
      }
    },
    "infant_span_ge37d": 192,
    "infant_span_ge37d_age_changed": 0,
    "expected_age_changes_if_age_at_collection": {
      "infant": 254.1,
      "child": 103.6,
      "adult": 49.1
    },
    "observed_age_changes": {
      "adult": 0,
      "child": 0,
      "infant": 0
    },
    "gap_days": {
      "n": 3180,
      "median": 16,
      "q1": 6,
      "q3": 33,
      "pct_0": 2.3,
      "pct_le1": 7.3,
      "pct_ge7": 74.8,
      "pct_ge28": 31.1,
      "max": 92
    }
  },
  "recorded_dates_vs_clinical_content": {
    "lag7_autocorrelation_detrended_daily_counts": 0.679,
    "clinics_with_50plus_tests": 33,
    "clinics_weekday_profile_beyond_chi2_95": 1,
    "chi2_6df_95th_percentile": 12.59,
    "weekday_only_clinics": {
      "radiation oncology": {
        "tests": 172,
        "two_lowest_weekday_pct": [
          9.9,
          12.2
        ]
      },
      "oncology day hosp": {
        "tests": 533,
        "two_lowest_weekday_pct": [
          12,
          12.8
        ]
      },
      "hem onc day hosp": {
        "tests": 71,
        "two_lowest_weekday_pct": [
          2.8,
          9.9
        ]
      }
    },
    "ed_vs_drive_through_weekday_chi2": 6.7,
    "nicu_recorded_age_q10_median_q90": [
      0.1,
      0.5,
      0.8
    ],
    "ct_time_r2_exceeds_perm_p95": false
  },
  "ct_association_time_vs_setting": {
    "calendar_time_only": {
      "r2": 0.008,
      "perm_p95": 0.0097,
      "perm_p99": 0.0132,
      "perm_share_ge_observed": 0.097
    },
    "setting_only": {
      "r2": 0.0502,
      "perm_p95": 0.0122,
      "perm_p99": 0.0172,
      "perm_share_ge_observed": 0
    }
  },
  "missing_values_by_stage": {
    "all_tests": {
      "age": 0,
      "pan_day": 0,
      "ct_result": 209,
      "clinic_name": 0,
      "patient_class": 7077,
      "demo_group": 0,
      "drive_thru_ind": 0
    },
    "positive": {
      "age": 0,
      "pan_day": 0,
      "ct_result": 5,
      "clinic_name": 0,
      "patient_class": 479,
      "demo_group": 0,
      "drive_thru_ind": 0
    },
    "quantifiable_ct": {
      "age": 0,
      "pan_day": 0,
      "ct_result": 0,
      "clinic_name": 0,
      "patient_class": 474,
      "demo_group": 0,
      "drive_thru_ind": 0
    },
    "age_le_100": {
      "age": 0,
      "pan_day": 0,
      "ct_result": 0,
      "clinic_name": 0,
      "patient_class": 474,
      "demo_group": 0,
      "drive_thru_ind": 0
    },
    "analytic": {
      "age": 0,
      "pan_day": 0,
      "ct_result": 0,
      "clinic_name": 0,
      "patient_class": 456,
      "demo_group": 0,
      "drive_thru_ind": 0
    },
    "analytic_patient_class_missing_by_setting": {
      "drive": 381,
      "outside": 74,
      "emergency": 0,
      "inpatient": 0,
      "outpatient": 1
    }
  },
  "flow_exclusions": {
    "negative_tests": 14358,
    "excluded_no_quantifiable_ct": 8,
    "excluded_age_gt_100": 1,
    "later_specimens_excluded": 71
  },
  "date_free_estimates": {
    "M1_unadjusted": {
      "infant": {
        "diff": -1.72,
        "lo": -3.65,
        "hi": 0.2
      },
      "child": {
        "diff": 0.97,
        "lo": -0.09,
        "hi": 2.03
      }
    },
    "setting_only": {
      "infant": {
        "diff": -2.11,
        "lo": -4.04,
        "hi": -0.18
      },
      "child": {
        "diff": 0.73,
        "lo": -0.39,
        "hi": 1.85
      }
    },
    "setting_only_fold_rna_infant": {
      "fold": 4.32,
      "lo": 1.13,
      "hi": 16.45
    },
    "setting_only_pr_ct_le24": {
      "pr": 1.51,
      "lo": 1.11,
      "hi": 2.06,
      "evalue_point": 2.39,
      "evalue_ci": 1.46
    },
    "setting_only_rv_point": 0.0771,
    "setting_only_rv_ci": 0.0097,
    "setting_only_tipping_or_infant": 1.2
  },
  "sensitivity_sample_sizes": {
    "no_outside_persons": 711,
    "no_outside_infants": 69,
    "no_outside_adults": 286,
    "all_specimens_persons": 785,
    "other_settings_infants": 57,
    "other_settings_adults": 63
  },
  "software": {
    "python": "3.12.13",
    "numpy": "2.5.2",
    "pandas": "3.0.5",
    "statsmodels": "0.15.0",
    "scipy": "1.18.1"
  },
  "setting_only_robustness": {
    "drive_through": {
      "diff": -2.64,
      "lo": -5.71,
      "hi": 0.43
    },
    "other_settings": {
      "diff": -1.59,
      "lo": -4.36,
      "hi": 1.19
    },
    "interaction_infant_x_drive": {
      "diff": -1.05,
      "lo": -5.19,
      "hi": 3.09
    },
    "all_specimens_cluster": {
      "diff": -1.77,
      "lo": -3.6,
      "hi": 0.06
    },
    "no_outside_institutions": {
      "diff": -2.46,
      "lo": -4.54,
      "hi": -0.38
    },
    "chop_patients_only": {
      "diff": -2.9,
      "lo": -5.36,
      "hi": -0.44
    },
    "selection_tilt": {
      "inf1.25_ad1": {
        "diff": 0.5,
        "lo": -1.54,
        "hi": 2.55
      },
      "inf1.5_ad1": {
        "diff": 2.54,
        "lo": 0.52,
        "hi": 4.57
      },
      "inf2_ad1": {
        "diff": 5.11,
        "lo": 3.27,
        "hi": 6.96
      },
      "inf1_ad1.5": {
        "diff": -5.66,
        "lo": -7.55,
        "hi": -3.76
      },
      "inf1_ad2": {
        "diff": -7.58,
        "lo": -9.45,
        "hi": -5.7
      }
    },
    "partial_r2_infant_ct": 0.0064,
    "partial_r2_setting_ct": 0.0511,
    "partial_r2_setting_infant": 0.1045,
    "clinics_beyond_chi2_95_expected_by_chance": 1.65
  }
}
Experiment script · experiment_v1.py
"""Age and SARS-CoV-2 RT-PCR cycle threshold (Ct) in deidentified CHOP testing records, 2020.
Cross-sectional analysis of each person's first quantifiable positive specimen. The records were
anonymized, time-shifted and permuted, so the script checks which age and date information survived,
estimates Ct differences versus adults (OLS, HC3 SEs) with and without the recorded-date spline and
testing setting, and adds sensitivity and bias analyses. Round 2 adds repeat-tester, weekday, split
Ct-integrity and missing-data checks and date-free models; round 3 repeats the robustness analyses
under the date-free setting-only model. Real data, loaded not generated."""
import json
import warnings

import numpy as np
import pandas as pd
import statsmodels.api as sm
import statsmodels.formula.api as smf
from scipy import stats

from dataset import load

warnings.filterwarnings("ignore")
rng = np.random.default_rng(20200401)
r = lambda x, k=2: float(round(float(x), k)) + 0.0  # + 0.0 turns -0.0 into 0.0

def setting_of(x):  # testing setting from the recorded location fields
    emerg = (x.patient_class == "emergency") | x.clinic_name.isin(["emergency dept", "urgent care center", "urg care ctr"])
    inpat = (x.patient_class.isin(["inpatient", "observation", "admit after surgery-ip", "admit after surgery-obs"])
             | x.clinic_name.str.contains("inpatient ward") | x.clinic_name.isin(["nicu", "picu"]))
    return np.select([x.drive_thru_ind == 1, x.demo_group == "client", emerg, inpat],
                     ["drive", "outside", "emergency", "inpatient"], "outpatient")

def rcs(x, t):  # restricted cubic spline basis (linear term + k-2 nonlinear terms)
    p, k = (lambda u: np.clip(u, 0, None) ** 3), len(t)
    return np.column_stack([x] + [(p(x - t[j]) - p(x - t[k - 2]) * (t[k - 1] - t[j]) / (t[k - 1] - t[k - 2])
                                   + p(x - t[k - 1]) * (t[k - 2] - t[j]) / (t[k - 1] - t[k - 2])) / (t[k - 1] - t[0]) ** 2
                                  for j in range(k - 2)])

def fit(formula, data, groups=None, weights=None):  # OLS/WLS with HC3, or cluster-robust by person
    if weights is not None:
        return smf.wls(formula, data, weights=weights).fit(cov_type="HC3")
    if groups is not None:
        return smf.ols(formula, data).fit(cov_type="cluster", cov_kwds={"groups": groups})
    return smf.ols(formula, data).fit(cov_type="HC3")

def est(m, term):
    lo, hi = m.conf_int().loc[term]
    return {"diff": r(m.params[term]), "lo": r(lo), "hi": r(hi)}

df = load()
print("loaded", df.shape, "| missing ct:", int(df.ct_result.isna().sum()), "| results:", df.result.value_counts().to_dict())

# ---- 1. Analytic sample: first quantifiable positive specimen per person ----
flow = {"tests": len(df), "persons": int(df.subject_id.nunique()), "invalid": int((df.result == "invalid").sum())}
d = df[df.result == "positive"]
flow["positive_tests"] = len(d)
d = d[d.ct_result.notna() & (d.ct_result < 45)]  # 45 is the not-detected placeholder
flow["quantifiable_positive_tests"] = len(d)
d = d[d.age <= 100]  # implausible recorded ages (119, 138) excluded
flow["after_age_le_100"] = len(d)
allpos = d.copy()
d = d.sort_values(["subject_id", "pan_day"], kind="mergesort").groupby("subject_id").head(1).copy()
flow["analytic_persons"] = len(d)
flow["persons_with_2plus_quantifiable_positives"] = int((allpos.subject_id.value_counts() > 1).sum())
KNOTS = np.percentile(d.pan_day, [5, 35, 65, 95])

def prepare(x):
    x = x.copy()
    x["ct"], x["setting"], x["week"] = x.ct_result, setting_of(x), x.pan_day // 7
    x["infant"], x["adult"] = (x.age < 1).astype(int), (x.age >= 18).astype(int)
    x["child"] = ((x.age >= 1) & (x.age < 18)).astype(int)
    x["a1_4"] = ((x.age >= 1) & (x.age < 5)).astype(int)
    x["a5_11"] = ((x.age >= 5) & (x.age < 12)).astype(int)
    x["a12_17"] = ((x.age >= 12) & (x.age < 18)).astype(int)
    x["u5"], x["a5_17"] = x.infant + x.a1_4, x.a5_11 + x.a12_17
    x["hvl"] = (x.ct <= 24).astype(int)  # high viral load marker (Ct <= 24)
    x[["s1", "s2", "s3"]] = rcs(x.pan_day.to_numpy(float), KNOTS)
    x["band"] = np.select([x.infant == 1, x.a1_4 == 1, x.a5_11 == 1, x.a12_17 == 1], ["infant", "1-4", "5-11", "12-17"], "adult")
    return x

d, allpos = prepare(d), prepare(allpos)
T, S = " + s1 + s2 + s3", " + C(setting, Treatment('drive'))"  # calendar time; setting (ref: drive-through)
PRES, COMP, FIVE = "ct ~ infant + child", "ct ~ u5 + a5_17", "ct ~ infant + a1_4 + a5_11 + a12_17"  # ref: adults
M3 = PRES + T + S  # primary model

# ---- 2. What survived deidentification? (all tests; one row per person at first test) ----
df["band3"] = np.select([df.age < 1, df.age < 18], ["infant", "child"], "adult")
first = df.groupby("subject_id").agg(age=("age", "first"), band=("band3", "first"), day1=("pan_day", "min"))
rho_age_day = {b: r(stats.spearmanr(g.age, g.day1)[0], 3) for b, g in first.groupby("band")}
third = pd.cut(first.day1, [3, 37, 72, 107], labels=["days4_37", "days38_72", "days73_107"])
age_by_third = {b: {k: r(v, 1) for k, v in g.groupby(third[g.index], observed=True).age.median().items()} for b, g in first.groupby("band")}
nicu, emp, obg = (df[df.clinic_name == c] for c in ("nicu", "employee health", "ob gyn"))
checks = {"nicu_tests": len(nicu), "nicu_pct_age_under1": r(100 * (nicu.age < 1).mean(), 1),
          "employee_health_tests": len(emp), "employee_health_pct_age18plus": r(100 * (emp.age >= 18).mean(), 1),
          "obgyn_tests": len(obg), "obgyn_pct_female": r(100 * (obg.gender == "female").mean(), 1)}
r2_age_time = r(smf.ols("age ~ s1 + s2 + s3", d[d.child == 1]).fit().rsquared, 3)

# Ct integrity: does Ct keep its association with setting and time beyond a permutation null?
Xp = np.column_stack([np.ones(len(d)), d[["s1", "s2", "s3"]], pd.get_dummies(d.setting, drop_first=True, dtype=float)])
def r2(y):
    e = y - Xp @ np.linalg.lstsq(Xp, y, rcond=None)[0]
    return 1 - e @ e / ((y - y.mean()) @ (y - y.mean()))
y = d.ct.to_numpy(float)
null = np.array([r2(rng.permutation(y)) for _ in range(1000)])
ct_integrity = {"r2_setting_time": r(r2(y), 3), "perm_null_p95": r(np.percentile(null, 95), 3),
                "perm_null_p99": r(np.percentile(null, 99), 3), "perm_share_ge_observed": r(np.mean(null >= r2(y)), 3)}
print("analytic n", len(d), "| spearman(age, first day):", rho_age_day, "| R2 age~time in 1-17:", r2_age_time)

# ---- 3. Description (Table 1) ----
def describe(g):
    q = g.ct.quantile([0.25, 0.5, 0.75])
    out = {"n": len(g), "ct_mean": r(g.ct.mean()), "ct_sd": r(g.ct.std()), "ct_median": r(q[0.5]), "ct_q1": r(q[0.25]),
           "ct_q3": r(q[0.75]), "pct_ct_le24": r(100 * g.hvl.mean(), 1), "day_median": r(g.pan_day.median(), 1),
           "day_q1": r(g.pan_day.quantile(0.25), 1), "day_q3": r(g.pan_day.quantile(0.75), 1)}
    out.update({f"pct_{s}": r(100 * (g.setting == s).mean(), 1) for s in ["drive", "emergency", "inpatient", "outpatient", "outside"]})
    return out

table1 = {b: describe(d[d.band == b]) for b in ["infant", "1-4", "5-11", "12-17", "adult"]}
table1["all"] = describe(d)

# ---- 4. Measures of association (mean Ct difference vs adults; HC3 95% CI) ----
assoc = {}
for name, base, terms in [("preserved", PRES, ["infant", "child"]), ("composite", COMP, ["u5", "a5_17"]),
                          ("recorded", FIVE, ["infant", "a1_4", "a5_11", "a12_17"])]:
    for model, rhs in [("M1", ""), ("M2", T), ("M3", T + S)]:
        m = fit(base + rhs, d)
        assoc[f"{name}_{model}"] = {t: est(m, t) for t in terms}
prim = assoc["preserved_M3"]["infant"]
fold = {"fold": r(2 ** (-prim["diff"])), "lo": r(2 ** (-prim["hi"])), "hi": r(2 ** (-prim["lo"]))}

# High viral load (Ct <= 24): prevalence ratio by modified Poisson regression (robust variance)
evalue = lambda x: float(x + np.sqrt(x * (x - 1))) if x >= 1 else 1.0
pr = {}
for name, base, term in [("infant_M1", "hvl ~ infant + child", "infant"), ("infant_M3", "hvl ~ infant + child" + T + S, "infant"),
                         ("u5_M3", "hvl ~ u5 + a5_17" + T + S, "u5")]:
    g = smf.glm(base, d, family=sm.families.Poisson()).fit(cov_type="HC0")
    b, (lo, hi) = np.exp(g.params[term]), np.exp(g.conf_int().loc[term])
    pr[name] = {"pr": r(b), "lo": r(lo), "hi": r(hi)}
    if name == "infant_M3":
        pr["evalue_point"], pr["evalue_ci"] = r(evalue(b)), r(evalue(lo))

# ---- 5. Effect-measure modification by setting and sensitivity analyses (infant vs adult) ----
d["drive"] = (d.setting == "drive").astype(int)
emm = {"drive_through": est(fit(PRES + T, d[d.drive == 1]), "infant"),
       "other_settings": est(fit(PRES + T + " + C(setting)", d[d.drive == 0]), "infant"),
       "n_infant_drive": int((d.drive * d.infant).sum()), "n_adult_drive": int((d.drive * d.adult).sum()),
       "interaction_infant_x_drive": est(fit(M3 + " + infant:drive + child:drive", d), "infant:drive")}
sens = {"weekly_time": est(fit(PRES + " + C(week)" + S, d), "infant"),
        "all_specimens_cluster": est(fit(M3, allpos, groups=allpos.subject_id), "infant"),
        "no_outside_institutions": est(fit(M3, d[d.setting != "outside"]), "infant"),
        "chop_patients_only": est(fit(M3, d[d.demo_group == "patient"]), "infant"),
        "n_all_specimens": len(allpos), "n_patients_only": int((d.demo_group == "patient").sum()),
        "n_adults_patients_only": int(((d.demo_group == "patient") & (d.adult == 1)).sum())}

# ---- 6. Quantitative bias analysis ----
# (a) unmeasured confounding: Cinelli-Hazlett robustness values (classical OLS t statistic)
mc = smf.ols(M3, d).fit()
t_, dof = mc.tvalues["infant"], mc.df_resid
f = abs(t_) / np.sqrt(dof)
fa = f - stats.t.ppf(0.975, dof - 1) / np.sqrt(dof - 1)
rv = lambda z: 0.5 * (np.sqrt(z ** 4 + 4 * z ** 2) - z ** 2) if z > 0 else 0.0
no_set = smf.ols(PRES + T, d).fit()
d_full, d_noset = smf.ols("infant ~ child" + T + S, d).fit(), smf.ols("infant ~ child" + T, d).fit()
conf = {"partial_r2_infant_ct": r(t_ ** 2 / (t_ ** 2 + dof), 4), "rv_point": r(rv(f), 4), "rv_ci": r(rv(fa), 4),
        "partial_r2_setting_ct": r((no_set.ssr - mc.ssr) / no_set.ssr, 4),
        "partial_r2_setting_infant": r((d_noset.ssr - d_full.ssr) / d_noset.ssr, 4)}

# (b) selection on viral load: inverse-probability-of-selection weights under an exponential-tilt model,
#     P(selected | Ct) proportional to OR ** ((30 - Ct) / 5) within infants or within adults.
def tilted(or_inf, or_ad, formula=M3):
    w = np.where(d.infant == 1, or_inf ** ((d.ct - 30) / 5), np.where(d.adult == 1, or_ad ** ((d.ct - 30) / 5), 1.0))
    return est(fit(formula, d, weights=w), "infant")

sel = {f"inf{a}_ad{b}": tilted(a, b) for a, b in [(1.25, 1), (1.5, 1), (2, 1), (1, 1.5), (1, 2)]}
lo_, hi_ = 0.0, np.log(20)
for _ in range(40):  # bisection for the infant selection OR that moves the estimate to zero
    mid = (lo_ + hi_) / 2
    lo_, hi_ = (mid, hi_) if tilted(np.exp(mid), 1)["diff"] < 0 else (lo_, mid)
sel["tipping_or_infant"] = r(np.exp(hi_))

# ======== Round 2 (referee): repeat testers, recorded dates vs clinical content, missing data, date-free models ========
# (7) Repeat testers: is recorded age the age at each collection, and do within-person gaps look like real intervals?
srt = df.sort_values(["subject_id", "pan_day"], kind="mergesort")
rep = srt[srt.subject_id.map(srt.subject_id.value_counts()) >= 2]
per = rep.groupby("subject_id").agg(band=("band3", "first"), ages=("age", "nunique"), age=("age", "first"),
                                     day1=("pan_day", "min"), dayn=("pan_day", "max"))
per["span"] = per.dayn - per.day1
gaps = rep.groupby("subject_id").pan_day.diff().dropna()
pi = per[per.band == "infant"]
expect = {"infant": (pi.span >= 36.525).sum() + (pi.span[pi.span < 36.525] / 36.525).sum()}  # tenths of a year must advance
expect.update({b: (np.minimum(per.span[per.band == b], 365.25) / 365.25).sum() for b in ["child", "adult"]})  # birthdays
repeat = {"persons": len(per), "tests": len(rep), "pct_constant_age": r(100 * (per.ages == 1).mean(), 1),
          "by_band": {b: {"persons": len(g), "pct_constant_age": r(100 * (g.ages == 1).mean(), 1), "span_median": r(g.span.median(), 1),
                          "rho_age_first_day": r(stats.spearmanr(g.age, g.day1)[0], 3), "rho_age_last_day": r(stats.spearmanr(g.age, g.dayn)[0], 3)}
                      for b, g in per.groupby("band")},
          "infant_span_ge37d": int((pi.span >= 37).sum()), "infant_span_ge37d_age_changed": int(((pi.span >= 37) & (pi.ages > 1)).sum()),
          "expected_age_changes_if_age_at_collection": {b: r(v, 1) for b, v in expect.items()},
          "observed_age_changes": {b: int((g.ages > 1).sum()) for b, g in per.groupby("band")},
          "gap_days": {"n": len(gaps), "median": r(gaps.median(), 1), "q1": r(gaps.quantile(0.25), 1), "q3": r(gaps.quantile(0.75), 1),
                       "pct_0": r(100 * (gaps == 0).mean(), 1), "pct_le1": r(100 * (gaps <= 1).mean(), 1),
                       "pct_ge7": r(100 * (gaps >= 7).mean(), 1), "pct_ge28": r(100 * (gaps >= 28).mean(), 1), "max": r(gaps.max(), 1)}}
# (8) Do recorded dates carry the records' clinical content? Weekly cycle, clinic weekday profiles, Ct vs date and setting.
day_n = df.groupby("pan_day").size().reindex(range(df.pan_day.min(), df.pan_day.max() + 1), fill_value=0).to_numpy(float)
dev = day_n - pd.Series(day_n).rolling(7, center=True, min_periods=4).mean().to_numpy()
dev = dev[~np.isnan(dev)]
p_all = (df.pan_day % 7).value_counts(normalize=True).sort_index().to_numpy()
crit = stats.chi2.ppf(0.95, 6)
def gof(x):
    o = (x.pan_day % 7).value_counts().reindex(range(7), fill_value=0).to_numpy(float)
    return float(((o - p_all * len(x)) ** 2 / (p_all * len(x))).sum())
clin = {c: gof(x) for c, x in df.groupby("clinic_name") if len(x) >= 50}
WEEKDAY = ["radiation oncology", "oncology day hosp", "hem onc day hosp"]
low2 = lambda x: [r(100 * v, 1) for v in sorted((x.pan_day % 7).value_counts(normalize=True).reindex(range(7), fill_value=0))[:2]]
ed_dt = df[(df.drive_thru_ind == 1) | ((df.clinic_name == "emergency dept") & (df.drive_thru_ind == 0))]
def r2x(X, yy):
    e = yy - X @ np.linalg.lstsq(X, yy, rcond=None)[0]
    return 1 - e @ e / ((yy - yy.mean()) @ (yy - yy.mean()))
def perm_r2(X):
    obs, nul = r2x(X, y), np.array([r2x(X, rng.permutation(y)) for _ in range(1000)])  # same generator, after the first null
    return {"r2": r(obs, 4), "perm_p95": r(np.percentile(nul, 95), 4), "perm_p99": r(np.percentile(nul, 99), 4), "perm_share_ge_observed": r(np.mean(nul >= obs), 3)}
one = np.ones((len(d), 1))
ct_split = {"calendar_time_only": perm_r2(np.column_stack([one, d[["s1", "s2", "s3"]]])),
            "setting_only": perm_r2(np.column_stack([one, pd.get_dummies(d.setting, drop_first=True, dtype=float)]))}
dates = {"lag7_autocorrelation_detrended_daily_counts": r(np.corrcoef(dev[:-7], dev[7:])[0, 1], 3),
         "clinics_with_50plus_tests": len(clin), "clinics_weekday_profile_beyond_chi2_95": int(sum(v > crit for v in clin.values())),
         "chi2_6df_95th_percentile": r(crit, 2),
         "weekday_only_clinics": {c: {"tests": int((df.clinic_name == c).sum()), "two_lowest_weekday_pct": low2(df[df.clinic_name == c])} for c in WEEKDAY},
         "ed_vs_drive_through_weekday_chi2": r(stats.chi2_contingency(pd.crosstab(ed_dt.drive_thru_ind, ed_dt.pan_day % 7))[0], 1),
         "nicu_recorded_age_q10_median_q90": [r(nicu.age.quantile(q), 1) for q in (0.1, 0.5, 0.9)],
         "ct_time_r2_exceeds_perm_p95": bool(ct_split["calendar_time_only"]["r2"] > ct_split["calendar_time_only"]["perm_p95"])}
# (9) Missing values per field at each stage of the sample flow; how records without patient class were classified
FIELDS = ["age", "pan_day", "ct_result", "clinic_name", "patient_class", "demo_group", "drive_thru_ind"]
st_pos = df[df.result == "positive"]
st_q = st_pos[st_pos.ct_result.notna() & (st_pos.ct_result < 45)]
stages = [("all_tests", df), ("positive", st_pos), ("quantifiable_ct", st_q), ("age_le_100", st_q[st_q.age <= 100]), ("analytic", d)]
missing = {k: {fl: int(v[fl].isna().sum()) for fl in FIELDS} for k, v in stages}
missing["analytic_patient_class_missing_by_setting"] = {sg: int((d.patient_class.isna() & (d.setting == sg)).sum())
                                                        for sg in ["drive", "outside", "emergency", "inpatient", "outpatient"]}
flow_x = {"negative_tests": int((df.result == "negative").sum()),
          "excluded_no_quantifiable_ct": flow["positive_tests"] - flow["quantifiable_positive_tests"],
          "excluded_age_gt_100": flow["quantifiable_positive_tests"] - flow["after_age_le_100"],
          "later_specimens_excluded": flow["after_age_le_100"] - flow["analytic_persons"]}
# (10) Estimates that do not use the recorded collection day: unadjusted (M1) and setting-only model
SO = PRES + S
so, mso = fit(SO, d), smf.ols(SO, d).fit()
ts, dofs = mso.tvalues["infant"], mso.df_resid
fs = abs(ts) / np.sqrt(dofs)
fsa = fs - stats.t.ppf(0.975, dofs - 1) / np.sqrt(dofs - 1)
gso = smf.glm("hvl ~ infant + child" + S, d, family=sm.families.Poisson()).fit(cov_type="HC0")
bso, (lso, hso) = np.exp(gso.params["infant"]), np.exp(gso.conf_int().loc["infant"])
lo_, hi_ = 0.0, np.log(20)
for _ in range(40):
    mid = (lo_ + hi_) / 2
    lo_, hi_ = (mid, hi_) if tilted(np.exp(mid), 1, SO)["diff"] < 0 else (lo_, mid)
eso = est(so, "infant")
date_free = {"M1_unadjusted": assoc["preserved_M1"], "setting_only": {t: est(so, t) for t in ["infant", "child"]},
             "setting_only_fold_rna_infant": {"fold": r(2 ** (-eso["diff"])), "lo": r(2 ** (-eso["hi"])), "hi": r(2 ** (-eso["lo"]))},
             "setting_only_pr_ct_le24": {"pr": r(bso), "lo": r(lso), "hi": r(hso), "evalue_point": r(evalue(bso)), "evalue_ci": r(evalue(lso))},
             "setting_only_rv_point": r(rv(fs), 4), "setting_only_rv_ci": r(rv(fsa), 4), "setting_only_tipping_or_infant": r(np.exp(hi_))}
sens_n = {"no_outside_persons": int((d.setting != "outside").sum()), "no_outside_infants": int(((d.setting != "outside") & (d.infant == 1)).sum()),
          "no_outside_adults": int(((d.setting != "outside") & (d.adult == 1)).sum()), "all_specimens_persons": int(allpos.subject_id.nunique()),
          "other_settings_infants": int(((d.drive == 0) & (d.infant == 1)).sum()), "other_settings_adults": int(((d.drive == 0) & (d.adult == 1)).sum())}
import sys, scipy, statsmodels
software = {"python": sys.version.split()[0], "numpy": np.__version__, "pandas": pd.__version__,
            "statsmodels": statsmodels.__version__, "scipy": scipy.__version__}
print("repeat testers", repeat["persons"], "| ct~time R2", ct_split["calendar_time_only"]["r2"], "| setting-only infant", date_free["setting_only"]["infant"])
# (11) Round 3 (referee): robustness analyses re-run under the setting-only specification (no recorded-date terms)
m_pres, i_c, i_cs = smf.ols(PRES, d).fit(), smf.ols("infant ~ child", d).fit(), smf.ols("infant ~ child" + S, d).fit()
so_rob = {"drive_through": est(fit(PRES, d[d.drive == 1]), "infant"),  # setting constant within the drive-through
          "other_settings": est(fit(PRES + " + C(setting)", d[d.drive == 0]), "infant"),
          "interaction_infant_x_drive": est(fit(SO + " + infant:drive + child:drive", d), "infant:drive"),
          "all_specimens_cluster": est(fit(SO, allpos, groups=allpos.subject_id), "infant"),
          "no_outside_institutions": est(fit(SO, d[d.setting != "outside"]), "infant"),
          "chop_patients_only": est(fit(SO, d[d.demo_group == "patient"]), "infant"),
          "selection_tilt": {f"inf{a}_ad{b}": tilted(a, b, SO) for a, b in [(1.25, 1), (1.5, 1), (2, 1), (1, 1.5), (1, 2)]},
          "partial_r2_infant_ct": r(ts ** 2 / (ts ** 2 + dofs), 4),
          "partial_r2_setting_ct": r((m_pres.ssr - mso.ssr) / m_pres.ssr, 4),
          "partial_r2_setting_infant": r((i_c.ssr - i_cs.ssr) / i_c.ssr, 4),
          "clinics_beyond_chi2_95_expected_by_chance": r(0.05 * len(clin), 2)}
print("setting-only robustness: drive", so_rob["drive_through"], "| other", so_rob["other_settings"])
res = {"design": "cross-sectional; first quantifiable positive specimen per person", "flow": flow,
       "deidentification": {"spearman_age_vs_first_test_day": rho_age_day, "median_recorded_age_by_first_test_period": age_by_third,
                            "persons_by_band": {b: int(n) for b, n in first.band.value_counts().items()},
                            "r2_recorded_age_on_time_children_1_17": r2_age_time,
                            "clinical_structure_checks": checks},
       "ct_integrity": ct_integrity, "spline_knots_day": [r(k, 1) for k in KNOTS], "table1": table1,
       "association_mean_ct_diff_vs_adults": assoc, "primary_infant_vs_adult_M3": prim, "fold_rna_infant_vs_adult": fold,
       "prevalence_ratio_ct_le24": pr, "effect_modification_setting": emm, "sensitivity_infant_vs_adult": sens,
       "bias_unmeasured_confounding": conf, "bias_selection_tilt": sel,
       "hypothesis_supported": bool(prim["hi"] < 0),
       "repeat_testers": repeat, "recorded_dates_vs_clinical_content": dates, "ct_association_time_vs_setting": ct_split,
       "missing_values_by_stage": missing, "flow_exclusions": flow_x, "date_free_estimates": date_free,
       "sensitivity_sample_sizes": sens_n, "software": software, "setting_only_robustness": so_rob}
print("RESULT_JSON: " + json.dumps(res))
Listing. experiment_v1.py

4.Analysis

Attempt 2Done

Analysis complete: experiment succeeded; inconclusive; 46 key numbers.

InconclusiveExperiment succeeded

Results and analysis

The revised script ran successfully on the first attempt and reproduced every previous estimate exactly. The 785 analytic persons are each represented by their first quantifiable positive specimen. The new checks change the interpretation. Among 1,744 persons tested 2 or more times, recorded age never changed. That held even for 192 infants whose tests spanned at least 37 days, against about 254, 104 and 49 expected changes in infants, children and adults if age were recorded at each collection. Within-person gaps looked like real retesting intervals (median 16 days), which rules out test-by-test re-sorting of dates. This favours ages reassigned per person in first-test order, but it cannot exclude whole per-person date sets re-placed by age. Further checks show that recorded dates are not tied to the records' clinical content either. Daily counts have a weekly cycle (lag-7 autocorrelation 0.679), yet only 1 of 33 clinics departs from the overall weekday profile, about what chance would give (1.65 expected), and clinics that ordinarily run on weekdays only show no weekend gap. Ct is associated with setting (R-squared 0.0502 vs permutation 95th percentile 0.0122) but not with recorded date (0.0080 vs 0.0097), although this check has limited power (3-df spline, 785 specimens). Calendar time therefore cannot be adjusted for, whichever field was reordered. Without the recorded dates, infants' Ct was 1.72 cycles lower than adults' unadjusted (95% CI -3.65 to 0.20) and 2.11 lower with setting adjustment (-4.04 to -0.18). The pre-specified model with the date spline gave -2.25 (-4.20 to -0.31). Prevalence ratios for Ct<=24 were 1.50 unadjusted and 1.51 adjusted. Children aged 1-17 y showed little difference (setting-only 0.73; -0.39 to 1.85), and ages 1-4 y remain unanswerable. The setting-only infant estimate is very fragile: robustness value for its interval 0.97%, against setting's date-free partial R-squared of 5.11% with Ct; tipping-point selection OR 1.20. Repeated without the spline, stratified and sensitivity estimates ranged from -1.59 to -2.90. The pre-specified flag is met, but because the question's calendar-time adjustment cannot be made and the date-free evidence is borderline and fragile, the verdict is inconclusive.

Key numbers

MetricValueInterpretation
Sample flow15524 tests from 12344 persons; 301 invalid; 14358 negative; 865 positive; 857 with quantifiable Ct (<45; 8 excluded); 856 after excluding recorded age >100 y (1 excluded); 785 persons in the analytic sample (71 later specimens excluded); 45 persons had 2+ quantifiable positive specimensCross-sectional sample of first quantifiable positive specimens.
Missing values by field and stageall_tests: ct_result 209, patient_class 7077; positive: ct_result 5, patient_class 479; quantifiable_ct: patient_class 474; age_le_100: patient_class 474; analytic: patient_class 456; analytic records without patient class by assigned setting: drive 381, outside 74, emergency 0, inpatient 0, outpatient 1Only Ct (missing in 5 positive tests, excluded) and patient class are ever missing. Patient class is missing for 456 analytic persons; setting was then assigned from the drive-through indicator (381), demographic group (74 outside institution) and clinic name. Only 1 fell to outpatient/other by default.
Spearman rho, recorded age vs first test day (persons, all tests)infants 0.801 (n=1500); children 1-17 y 0.975 (n=7179); adults 0.644 (n=3665)Recorded age is ordered by first test date within every band, so within-band age is unusable in all three bands.
Median recorded age by first-test period (days 4-37 / 38-72 / 73-107)infants 0.1 / 0.4 / 0.8 y; children 1-17 y 2.0 / 6.0 / 14.0 y; adults 25.0 / 33.0 / 47.0 yNo testing programme produces this. It is a deidentification artefact.
R-squared of recorded age on recorded-day spline, analytic children 1-17 y0.671Two thirds of recorded-age variance among analysed children follows the recorded day.
Checks against clinical structureNICU tests recorded <1 y: 97.6% of 294; employee-health tests recorded >=18 y: 100.0% of 12; OB-GYN tests recorded female: 56.5% of 69; NICU recorded age 10th/50th/90th percentile 0.1/0.5/0.8 yBand membership is supported (two checks, 294 and 12 tests). Recorded sex is not credible. NICU recorded ages are spread across the first year rather than concentrated in the neonatal period, so ages within the infant band are detached as well.
Repeat testers (2+ tests of any result)1744 persons, 4924 tests; recorded age constant within person in 100.0% (infants 323, children 938, adults 483; all 100.0%); infants whose tests span >=37 days: 192, of whom 0 show any age change; expected age changes if age were recorded at each collection: infants 254.1, children 103.6, adults 49.1; observed 0, 0, 0Recorded age is a per-person value, not the age at each collection. In 192 infants, age in tenths of a year would have had to advance.
Within-person gaps between consecutive tests (days)n=3180; median 16.0 (IQR 6.0-33.0); same day 2.3%; <=1 day 7.3%; >=7 days 74.8%; >=28 days 31.1%; max 92.0These look like real retesting intervals. If dates had been re-sorted test by test by age, a person's tests would sit on adjacent dates.
Repeat testers: Spearman rho of recorded age with first vs last test dayinfants 0.431 vs 0.458; children 0.903 vs 0.276; adults 0.551 vs 0.149The ordering is anchored on each person's first test. This fits ages reassigned in first-test order, or whole per-person date sets re-placed by age; the repeat-tester check cannot separate those two.
Recorded dates vs clinical content (weekday structure)lag-7 autocorrelation of detrended daily test counts 0.679; clinics (>=50 tests) whose weekday profile departs from the overall profile beyond the chi-squared(6) 95th percentile (12.59): 1 of 33; two lowest weekday shares: radiation oncology (172 tests) 9.9% and 12.2%, oncology day hospital (533) 12.0% and 12.8%, hematology-oncology day hospital (71) 2.8% and 9.9%; emergency department vs drive-through weekday chi-squared 6.7 (95th percentile 12.59)The recorded dates carry a real weekly cycle overall, but only 1 of 33 clinics departs from the overall weekday profile, about what chance alone would give (1.65 expected at the 95% level). Clinics that ordinarily run on weekdays only also show no weekend gap, which mainly carries the detachment inference. The emergency department vs drive-through comparison is uninformative, because both operate daily.
Ct integrity split (analytic sample): R-squared vs 1,000-permutation nullrecorded-day spline only 0.0080 (null 95th 0.0097, 99th 0.0132; share of permutations >= observed 0.097); setting only 0.0502 (null 95th 0.0122, 99th 0.0172; share 0.0); combined 0.055 (95th 0.018, 99th 0.022)Ct keeps its association with setting but shows none beyond chance with the recorded date. The earlier combined check was carried by setting. Power against a modest epidemic-phase trend is limited (3-df spline, 785 specimens).
Calendar-time spline knots (recorded collection day)19.0, 47.0, 78.0, 103.8Restricted cubic spline, 4 knots at the 5th/35th/65th/95th percentiles (3 df).
Table 1 - All 785n=785; Ct mean 28.97 (SD 7.29); median 28.36 (IQR 22.60-35.87); Ct<=24 32.0%; recorded collection day median 63.0 (IQR 38.0-86.0); setting % drive-through 57.6, emergency 20.6, inpatient 4.6, outpatient/other 7.8, outside institution 9.4Descriptive characteristics of first quantifiable positive specimens.
Table 1 - Infants <1 yn=79; Ct mean 26.94 (SD 7.96); median 24.32 (IQR 20.60-34.68); Ct<=24 46.8%; recorded collection day median 52.0 (IQR 30.0-71.0); setting % drive-through 27.8, emergency 38.0, inpatient 6.3, outpatient/other 15.2, outside institution 12.7Descriptive characteristics of first quantifiable positive specimens.
Table 1 - Recorded 1-4 yn=107; Ct mean 30.85 (SD 7.74); median 33.56 (IQR 23.31-37.80); Ct<=24 28.0%; recorded collection day median 29.0 (IQR 23.0-35.0); setting % drive-through 42.1, emergency 29.0, inpatient 15.9, outpatient/other 13.1, outside institution 0.0Descriptive characteristics of first quantifiable positive specimens.
Table 1 - Recorded 5-11 yn=117; Ct mean 28.92 (SD 7.60); median 28.19 (IQR 22.18-36.20); Ct<=24 34.2%; recorded collection day median 53.0 (IQR 44.0-63.0); setting % drive-through 49.6, emergency 33.3, inpatient 2.6, outpatient/other 12.8, outside institution 1.7Descriptive characteristics of first quantifiable positive specimens.
Table 1 - Recorded 12-17 yn=168; Ct mean 29.36 (SD 6.91); median 28.57 (IQR 23.51-36.09); Ct<=24 27.4%; recorded collection day median 84.5 (IQR 72.8-97.0); setting % drive-through 45.2, emergency 26.2, inpatient 2.4, outpatient/other 6.0, outside institution 20.2Descriptive characteristics of first quantifiable positive specimens.
Table 1 - Adults >=18 yn=314; Ct mean 28.66 (SD 6.91); median 27.93 (IQR 22.74-35.02); Ct<=24 31.2%; recorded collection day median 70.5 (IQR 43.2-91.0); setting % drive-through 79.9, emergency 5.7, inpatient 2.2, outpatient/other 3.2, outside institution 8.9Descriptive characteristics of first quantifiable positive specimens.
DATE-FREE: infants vs adults, unadjusted (M1)-1.72 (95% CI -3.65 to 0.20)No use of recorded dates. The interval includes 0.
DATE-FREE: infants vs adults, setting-only model-2.11 (95% CI -4.04 to -0.18); children 1-17 y 0.73 (95% CI -0.39 to 1.85)Adjusted for testing setting, without the recorded-date spline. Only the setting-adjusted interval excludes 0, and narrowly.
Fold difference in RNA, infants vs adults, setting-only4.32 (95% CI 1.13 to 16.45)Assumes 100% amplification efficiency (one doubling per cycle).
Prevalence ratio of Ct<=24, infants vs adults, setting-only1.51 (1.11 to 2.06); E-value 2.39 (confidence limit 1.46)Date-free. The setting-adjusted PR matches the M3 PR.
Setting-only model bias analysisrobustness value 0.0771 (estimate to 0), 0.0097 (conventional-SE interval to include 0); tipping-point infant selection OR per 5 cycles lower Ct 1.20Very fragile. A confounder explaining under 1% of residual variance, or modest differential selection, removes the date-free adjusted difference.
Preserved bands M1: mean Ct difference vs adultsinfants -1.72 (95% CI -3.65 to 0.20); children 1-17 y 0.97 (95% CI -0.09 to 2.03)Unadjusted.
Preserved bands M2: mean Ct difference vs adultsinfants -1.94 (95% CI -3.88 to 0.00); children 1-17 y 0.93 (95% CI -0.13 to 1.99)Adjusted for the recorded-date spline (not interpretable as calendar time).
Preserved bands M3: mean Ct difference vs adultsinfants -2.25 (95% CI -4.20 to -0.31); children 1-17 y 0.72 (95% CI -0.40 to 1.85)Recorded-date spline and setting; the pre-specified PRIMARY model, whose calendar-time premise the new checks undermine.
Composite coding M1: mean Ct difference vs adultsunder 5 (infants + recorded 1-4 y) 0.53 (95% CI -0.86 to 1.92); 5-17 y 0.52 (95% CI -0.62 to 1.65)Not interpretable as an age contrast.
Composite coding M2: mean Ct difference vs adultsunder 5 (infants + recorded 1-4 y) -0.13 (95% CI -1.60 to 1.34); 5-17 y 0.80 (95% CI -0.35 to 1.95)Not interpretable as an age contrast.
Composite coding M3: mean Ct difference vs adultsunder 5 (infants + recorded 1-4 y) -0.52 (95% CI -2.06 to 1.03); 5-17 y 0.74 (95% CI -0.46 to 1.94)Not interpretable as an age contrast.
Recorded five bands M1: mean Ct difference vs adultsinfants -1.72 (95% CI -3.65 to 0.20); 1-4 y 2.19 (95% CI 0.53 to 3.85); 5-11 y 0.26 (95% CI -1.32 to 1.84); 12-17 y 0.70 (95% CI -0.60 to 1.99)Sub-bands within childhood are not interpretable as age groups.
Recorded five bands M2: mean Ct difference vs adultsinfants -1.92 (95% CI -3.87 to 0.02); 1-4 y 1.53 (95% CI -0.31 to 3.36); 5-11 y 0.36 (95% CI -1.34 to 2.05); 12-17 y 0.94 (95% CI -0.41 to 2.29)Sub-bands within childhood are not interpretable as age groups.
Recorded five bands M3: mean Ct difference vs adultsinfants -2.20 (95% CI -4.16 to -0.24); 1-4 y 1.13 (95% CI -0.78 to 3.05); 5-11 y 0.75 (95% CI -0.99 to 2.49); 12-17 y 0.51 (95% CI -0.89 to 1.91)Sub-bands within childhood are not interpretable as age groups.
Pre-specified PRIMARY: infants vs adults, M3-2.25 (95% CI -4.20 to -0.31)The pre-specified criterion (upper limit <0) is met, but the model's recorded-date spline cannot be read as calendar time. The estimate is close to the date-free setting-only estimate.
Fold difference in RNA, infants vs adults, M34.76 (95% CI 1.24 to 18.38)Assumes 100% efficiency.
Prevalence ratio of Ct<=24, infants vs adultsM1 (unadjusted, date-free) 1.50 (1.13 to 2.00); M3 1.51 (1.11 to 2.06); E-value for M3 2.39 (limit 1.46)Even unadjusted, infants more often had Ct<=24.
Prevalence ratio of Ct<=24, under 5 (composite) vs adults, M31.27 (0.95 to 1.70)Not interpretable as an age contrast.
Effect-measure modification by setting (infants vs adults)drive-through -2.73 (95% CI -5.77 to 0.30) [22 infants, 251 adults]; other settings -1.90 (95% CI -4.79 to 0.99) [57 infants, 63 adults]; infant x drive-through product term -1.08 (95% CI -5.22 to 3.07)Both stratum models include the recorded-date spline. Directions agree and both are imprecise.
Sensitivity analyses (infants vs adults, M3 specification)weekly indicators -2.25 (95% CI -4.20 to -0.30); all 856 quantifiable positive specimens from 785 persons with person-clustered SEs -1.79 (95% CI -3.63 to 0.06); excluding outside institutions (711 persons: 69 infants, 286 adults) -2.56 (95% CI -4.66 to -0.45); hospital patients only (n=484, 76 adults) -2.83 (95% CI -5.32 to -0.33)All use the recorded-date terms. Directions agree.
Unmeasured confounding for M3 (Cinelli-Hazlett, conventional OLS t)partial R2 of infant status with Ct 0.0072; robustness value 0.0818 (estimate to 0), 0.0147 (conventional-SE interval to include 0); setting partial R2 with Ct 0.0488, with infant status 0.1022Fragile. A confounder weaker than setting suffices.
Selection on viral load (exponential-tilt IPSW), M3infant selection OR per 5-cycle lower Ct 1.25: 0.32 (95% CI -1.72 to 2.37); 1.5: 2.33 (95% CI 0.31 to 4.34); 2: 4.84 (95% CI 3.00 to 6.68); adult OR 1.5: -5.74 (95% CI -7.66 to -3.83); adult OR 2: -7.63 (95% CI -9.52 to -5.75); tipping-point infant OR 1.22Differential selection of modest size in either direction changes the answer.
Setting-only specification: effect-measure modification by setting (infants vs adults)drive-through -2.64 (95% CI -5.71 to 0.43) [22 infants, 251 adults]; other settings -1.59 (95% CI -4.36 to 1.19) [57 infants, 63 adults]; infant x drive-through product term -1.05 (95% CI -5.19 to 3.09)Date-free versions of the stratified analyses. Directions agree, both strata are imprecise, and there is no evidence of modification.
Setting-only specification: sensitivity analyses (infants vs adults)all 856 quantifiable positive specimens from 785 persons with person-clustered SEs -1.77 (95% CI -3.60 to 0.06); excluding outside institutions (711 persons: 69 infants, 286 adults) -2.46 (95% CI -4.54 to -0.38); hospital patients only (n=484, 76 adults) -2.90 (95% CI -5.36 to -0.44)Date-free versions. Directions agree with the setting-only estimate, and the all-specimen interval includes 0. The weekly-indicator analysis (M3 specification) is only a check that the time coding is immaterial.
Setting-only specification: benchmarks for the robustness valuespartial R2 of infant status with Ct 0.0064; setting with Ct 0.0511; setting with infant status 0.1045Date-free benchmarks. Setting explains far more than the 0.97% a confounder would need to move the setting-only conventional-SE interval to include 0.
Setting-only specification: selection on viral load (exponential-tilt IPSW)infant selection OR per 5-cycle lower Ct 1.25: 0.50 (95% CI -1.54 to 2.55); 1.5: 2.54 (95% CI 0.52 to 4.57); 2: 5.11 (95% CI 3.27 to 6.96); adult OR 1.5: -5.66 (95% CI -7.55 to -3.76); adult OR 2: -7.58 (95% CI -9.45 to -5.70); tipping-point infant OR 1.20Date-free versions. Modest differential selection in either direction changes the answer.
Softwarepython 3.12.13, numpy 2.5.2, pandas 3.0.5, statsmodels 0.15.0, scipy 1.18.1Versions used for the executed analysis.
hypothesis_supported (script flag, pre-specified M3 rule)trueThe flag reports the pre-specified rule. The rule's premise, that the recorded day is calendar time, is not supported, so the overall verdict is inconclusive.

Caveats

  • Within every band, recorded age is ordered by first test date (rho 0.801, 0.975 and 0.644). Within-band age is unusable in all three bands. 'Preserved' refers to band membership only, supported by two checks (NICU 294 tests; employee health 12 tests). The rho >= 0.9 decision threshold was arbitrary.
  • The repeat-tester check shows that recorded age is a per-person value and that dates were not re-sorted test by test. It cannot separate reassigned ages from whole per-person date sets re-placed by age.
  • Recorded dates appear detached from the records' clinical content. Only 1 of 33 clinics departs from the overall weekday profile, about what chance would give, and weekday-only clinics show no weekend gap; this mainly carries the inference, which assumes those clinics ordinarily operate on weekdays only. The emergency department vs drive-through comparison is uninformative because both operate daily. Ct shows no association with recorded date, but that check has limited power (3-df spline, 785 specimens). Calendar time, and hence the question's calendar-time adjustment, cannot be addressed, and M2/M3 spline terms are not calendar-time adjustments.
  • Date-free infant estimates are borderline: unadjusted -1.72 (-3.65 to 0.20) includes 0, and setting-only -2.11 (-4.04 to -0.18) is very fragile (robustness value for the interval 0.97%; tipping-point selection OR 1.20).
  • Selection: Ct is observed only in people tested and positive, and modest differential selection on viral load could create or remove the infant difference.
  • Recorded sex was altered (OB-GYN 56.5% female) and not used. Patient class is missing for 456 analytic persons; setting was assigned from other fields, with only 1 falling to outpatient/other by default.
  • Measurement: assay platform and specimen type are not recorded. The Ct<=24 threshold comes from an E-gene culture study, and the fold conversion assumes 100% efficiency.
  • Precision: 79 infants. Effect-modification, sensitivity and selection analyses are reported under both the setting-only and M3 specifications.
  • The adult comparison group is unusual: 79.9% were tested at the drive-through and only 76 were hospital patients. The study is single-centre, from 2020, with time-shifted dates. The analysis plan was fixed after the data structure was inspected and was not preregistered. The setting-only model and weekday checks were added in revision, prompted by the identification problem.

5.Paper writing

Attempt 8Done

Manuscript drafted: “Infant–Adult Differences in SARS-CoV-2 Cycle Threshold Values and the Limits of Deidentified Testing Records: A Cross-Sectional Analysis” (≈ 6075 words).

Reviewer feedback applied: Independent referee (Claude Fable 5.1), round 2: minor_revision. Required changes: (1) [analysis] Re-run the within-setting contrasts, the infant x drive-through product term, the three applicable sensitivity analyses (all quantifiable specimens with person-clustered SEs, excluding outside institutions, hospital patients only) and the selection-OR scenarios under the setting-only specification, and compute setting's partial R-squared with Ct and with infant status without the spline as the benchmark for the setting-only robustness values; report these in Table 4 in place of, or beside, the M3 versions, and label the weekly-indicator row as a check that the time coding is immaterial. (2) [text] In Results and the Abstract, state the weekday evidence as 1 of 33 clinics departing from the overall profile (about what chance would give), drop or explicitly discount the emergency-department vs drive-through comparison since both operate daily, and add that the Ct-date check has limited power (3-df spline, 785 specimens) so the detachment conclusion rests mainly on the uniform profiles of weekday-only clinics. (3) [text] Qualify the Yonker citation in the Introduction: no age gradient within that pediatric cohort, but early viral loads above those of hospitalized adults, so cite it for the former rather than as a no-difference study.

≈ 6075 words · AI-drafted manuscript · not peer-reviewed

Infant–Adult Differences in SARS-CoV-2 Cycle Threshold Values and the Limits of Deidentified Testing Records: A Cross-Sectional Analysis

Abstract

Young children are often said to carry more severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) than adults, but comparisons rarely consider why and when people were tested. In a cross-sectional analysis of 15,524 deidentified reverse-transcription polymerase chain reaction tests from the Children's Hospital of Philadelphia (2020), we studied each person's first positive specimen with a quantifiable cycle threshold, or Ct (785 persons). Lower Ct means more RNA. Within age bands, recorded age was ordered by date of first test (Spearman ρ = 0.975 at ages 1–17 years) and never changed within 1,744 repeat testers. Only 1 of 33 clinics had its own weekday pattern, about what chance would give, and recorded dates showed no association with Ct, so neither ages 1–4 years nor calendar time could be analyzed. Without recorded dates, infants' Ct was 1.72 cycles lower than adults' unadjusted (95% confidence interval: −3.65, 0.20) and 2.11 cycles lower with setting adjustment (95% confidence interval: −4.04, −0.18), about 4.3-fold more RNA. Infant-specific selection on viral load (odds ratio 1.20 per 5 cycles) would remove it. These records are compatible with a small infant excess but cannot show whether children under 5 years differ from adults once calendar time is considered.

Keywords: SARS-CoV-2; cycle threshold; viral load; infants; selection bias; quantitative bias analysis; deidentification; cross-sectional studies

Introduction

Whether young children carry more virus than adults matters for child care, schools and households. In Ontario, pediatric index cases aged 0–3 years had the highest odds of household transmission of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) (1). Upper-respiratory viral RNA, usually measured by the reverse-transcription polymerase chain reaction (RT-PCR) cycle threshold (Ct), is the most widely available proxy for infectiousness. Lower Ct values mean more RNA.

The evidence is contradictory. A study of 145 symptomatic patients at a Chicago children's hospital reported median Ct values of 6.5 in children younger than 5 years and 11.0 in adults, read by its authors as a 10- to 100-fold excess (2). Much larger laboratory series disagree. In the Netherlands viral load increased with age (3), and in Germany the youngest subjects had loads at most 0.5 log10 units lower than older subjects (4). Multicenter and Geneva studies found no difference between children and adults (5, 6). A Boston pediatric cohort showed no age gradient in viral load, although children's early loads exceeded those of hospitalized adults (7), and in Manitoba children had higher Ct than adults (8). In community surveillance, symptomatic children and symptomatic adults differed by −0.7 cycles (95% confidence interval (CI): −2.2, 0.9) (9). Infant-specific studies point to heterogeneity within the under-5 group: the highest loads were in infants younger than 6 months (10) and in hospitalized infants younger than 1 year (11).

Two features of testing may explain much of this disagreement. The first is indication. Asymptomatic children's median adjusted Ct was 10.3 cycles higher than symptomatic children's in 9 pediatric hospital programs (12). At the Children's Hospital of Philadelphia (CHOP), criteria moved within weeks from symptomatic testing to preprocedural and universal preadmission screening (13). The second is calendar time: the Ct distribution among positive specimens shifts with epidemic phase (14, 15). Ct is observed only in people who tested positive, so selection on symptoms, and hence on viral load, creates a rarely quantified collider structure (16, 17). At a pediatric hospital the adult comparison group is unusual too. CHOP also tested parents, household contacts, exposed health care workers and pregnant women (13), and it processed specimens referred from outside institutions.

We asked whether children younger than 5 years had lower Ct values than adults once testing setting and calendar time were accounted for. The data are an openly distributed teaching dataset of CHOP records, documented only as "anonymized, time-shifted, and permuted" (18). Disclosure control by swapping can rearrange joint distributions (19). We therefore first established which age and date information survived deidentification, then estimated age contrasts with quantitative bias analysis.

Methods

Study Population

This is a cross-sectional analysis of routinely collected laboratory records, reported following the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement and its RECORD extension for routinely collected health data (20, 21). The data are the covid_testing dataset of the R package medicaldata, version 0.2.0, retrieved from the Rdatasets archive on October 24, 2026 (22). The file holds 15,524 deidentified SARS-CoV-2 RT-PCR tests from 12,344 persons at CHOP, a pediatric hospital in Philadelphia, Pennsylvania, recorded on days 4–107 of the 2020 pandemic period on a time-shifted scale. The data were donated by Amrom E. Obstfeld, and there is no source publication (18).

We kept positive tests with a quantifiable Ct, that is, non-missing and below 45, the value carried by negative results. We excluded recorded ages above 100 years and kept each person's first quantifiable positive specimen. Figure 1 gives the flow and the missing values at each stage.

The assumed causal diagram, in words, is as follows. Age may affect Ct through viral replication and shedding. Testing setting depends on age and illness, and shapes the Ct distribution, because screening finds late or low-load infections (12). Calendar time drives testing policy (13) and epidemic phase (14). Inclusion requires testing positive, which depends on age, setting, symptoms and viral load, so restricting to positives conditions on a collider (16, 17). Time since infection, symptoms, specimen type and assay are unmeasured.

Exposure Assessment

Recorded age was in years, with infants in tenths of a year. We formed 5 bands: infants (<1 year), 1–4, 5–11 and 12–17 years, and adults (≥18 years, the reference). We also formed a composite under-5 group (infants plus 1–4 years). We checked what survived deidentification:

  1. Age against first test. The Spearman correlation between recorded age and first test day within each band.
  2. Clinical structure. Whether neonatal intensive care unit (NICU) tests were recorded as younger than 1 year, employee-health tests as 18 years or older, and obstetrics and gynecology (OB-GYN) tests as female.
  3. Repeat testers. Among persons tested 2 or more times, whether recorded age is constant within person, against the changes expected if age were recorded at each collection: infants whose tests span at least 37 days, whose age in tenths must advance, plus birthdays within each span. Also the within-person gaps between tests, and the correlation of age with first and last test day.
  4. Recorded dates against clinical content. The weekly cycle in daily counts, and whether each clinic's weekday profile departs from the overall profile, judged by a χ² statistic on 6 degrees of freedom (df), with attention to clinics that ordinarily operate on weekdays only.
  5. Ct against setting and against recorded date. Each was assessed separately and jointly, by the coefficient of determination (R²) against 1,000 permutations of Ct.

Checks 3 and 4 and the separate Ct checks were added at peer review. They ask whether ages were reordered by date or dates by age, and whether recorded dates carry the records' clinical content at all. We set a decision rule after inspecting the data, with an arbitrary threshold: where recorded age was near-monotone in first test day within a band (ρ ≥ 0.9), sub-bands of that band would not be interpreted as age groups.

Outcome Ascertainment

The outcome was the RT-PCR Ct value, in cycles. Ct depends on assay, gene target and specimen collection (23). The platform is not recorded; in 2020 CHOP used a laboratory-developed N2 assay and the Xpert Xpress test (13). The secondary outcome was Ct ≤24, a marker of likely culturable virus taken from an E-gene culture study (24). Mean differences (Δ\Delta) were translated into fold differences in RNA as 2−Δ2^{-\Delta}, which assumes 100% amplification efficiency.

Covariates

The recorded collection day was modeled as a restricted cubic spline with knots at the 5th, 35th, 65th and 95th percentiles (days 19.0, 47.0, 78.0 and 103.8). Testing setting was assigned in this order of precedence:

  1. drive-through (drive-through indicator equal to 1);
  2. outside institution (specimens referred by client institutions);
  3. emergency department or urgent care;
  4. inpatient (inpatient, observation or post-surgical class, or an inpatient-ward, NICU or pediatric intensive care unit location);
  5. outpatient or other.

Patient class was missing for 456 analytic persons. Their setting came from the drive-through indicator (381) or demographic group (74, outside institution); clinic name placed none in emergency or inpatient care, and 1 person fell to outpatient/other by default. Sex, which was not credible, and payer were not used.

Statistical Analysis

Mean differences in Ct relative to adults were estimated by ordinary least squares with heteroskedasticity-consistent (HC3) standard errors (25). Four models were fitted:

  • M1, unadjusted;
  • a setting-only model, added at review, which does not use recorded dates;
  • M2, adjusted for the recorded-day spline;
  • M3, adjusted for the spline and setting, the pre-specified primary model:
Cti=β0+β1 Infanti+β2 Childi+f(dayi)+∑kγk Settingik+εi,\mathrm{Ct}_i=\beta_0+\beta_1\,\mathrm{Infant}_i+\beta_2\,\mathrm{Child}_i+f(\mathrm{day}_i)+\textstyle\sum_k\gamma_k\,\mathrm{Setting}_{ik}+\varepsilon_i,

where Child denotes ages 1–17 years and the drive-through is the reference setting. M1–M3 were also fitted to the composite coding (under 5; 5–17 years) and to the 5-band coding. Covariate coefficients are not interpreted (26). For Ct ≤24, prevalence ratios (PR) were estimated by modified Poisson regression with robust variance (27), with E-values (28).

Effect-measure modification was examined within the drive-through and within the other settings, and with infant × drive-through and child × drive-through product terms. Sensitivity analyses used all quantifiable positive specimens with person-clustered standard errors, excluded outside institutions, and restricted to hospital patients. All were run under both the setting-only and the M3 specifications. Replacing the spline with weekly indicators, in M3 only, checks that the time coding is immaterial.

Quantitative bias analysis had 2 parts.

  • Unmeasured confounding. For the infant coefficient of the setting-only model and of M3, we computed Cinelli–Hazlett robustness values from the conventional ordinary least squares t statistic (29). They give the share of residual variance in both infant status and Ct that a confounder would need to explain to move the estimate to 0, or its conventional-standard-error interval to include 0. We benchmarked them against the partial R² of setting in the same model.
  • Selection on viral load. Selection was modeled with exponential-tilt inverse-probability-of-selection weights (30–33), in which the probability of being tested and detected changes by an odds ratio (OR) for every 5 cycles lower Ct:
wi=ORg(i)(Cti−30)/5,w_i=\mathrm{OR}_{g(i)}^{(\mathrm{Ct}_i-30)/5},

where g(i)g(i) is the person's band, persons aged 1–17 years received weight 1, and weights were centered at Ct 30. In both models, the scenarios set the OR to 1.25, 1.5 or 2 for infants (adults 1), or to 1.5 or 2 for adults (infants 1). For both M3 and the setting-only model, the infant OR at which the difference reaches 0 was found by bisection over ORs from 1 to 20. The range reflects the cited evidence. Symptomatic people had Ct values about 3 cycles lower than asymptomatic people, and asymptomatic infection was commoner in infected children than adults (38.2% vs 7.2%) (9); asymptomatic children's median Ct was 10.3 cycles higher (12). The tipping point depends on the tilt form, centering and weights.

The permutation nulls used a single generator with seed 20200401: 1,000 permutations for the joint check, then 1,000 each for setting and for recorded date. Analyses used Python 3.12.13 with NumPy 2.5.2, pandas 3.0.5, statsmodels 0.15.0 and SciPy 1.18.1. The analysis plan was fixed after the data structure was inspected and was not preregistered.

Results

Of 15,524 tests, 865 were positive and 785 persons contributed a first quantifiable positive specimen (Figure 1): 79 infants; 107, 117 and 168 persons with recorded ages of 1–4, 5–11 and 12–17 years; and 314 adults (Table 1). Mean Ct was 28.97 (standard deviation (SD), 7.29) and median Ct 28.36 (interquartile range (IQR), 22.60–35.87); 32.0% of specimens had a Ct ≤24. Of adults, 79.9% were tested at the drive-through, against 27.8% of infants, 38.0% of whom were tested in the emergency department or urgent care.

Figure 1. Flow of Tests and Persons, With Missing Values at Each Stage, Children's Hospital of Philadelphia Testing Records, 2020

StageTests or personsExcluded before the next stageMissing values
All recorded RT-PCR tests15,524 tests (12,344 persons)301 invalid; 14,358 negativeCt 209; patient class 7,077
↓ Positive results865 tests8 without a quantifiable CtCt 5; patient class 479
↓ Quantifiable Ct (<45)857 tests1 with recorded age >100 yearspatient class 474
↓ Recorded age ≤100 years856 tests71 later specimens of the same personspatient class 474
↓ First quantifiable positive specimen per person785 persons—patient class 456

Abbreviations: Ct, cycle threshold; RT-PCR, reverse-transcription polymerase chain reaction. Note: No values were missing for age, collection day, clinic name, demographic group or drive-through indicator at any stage. Setting assignment for missing patient class is described in Methods.

Table 1. Characteristics of First Quantifiable Positive Specimens by Recorded Age Band, Children's Hospital of Philadelphia, 2020 (n = 785)

CharacteristicInfants, <1 y1–4 y†5–11 y†12–17 y†Adults, ≥18 yAll
No. of persons79107117168314785
Drive-through, %27.842.149.645.279.957.6
Emergency/urgent care, %38.029.033.326.25.720.6
Inpatient, %6.315.92.62.42.24.6
Outpatient/other, %15.213.112.86.03.27.8
Outside institution, %12.70.01.720.28.99.4
Recorded collection day, median (IQR)52.0 (30.0–71.0)29.0 (23.0–35.0)53.0 (44.0–63.0)84.5 (72.8–97.0)70.5 (43.2–91.0)63.0 (38.0–86.0)
Ct, mean (SD)26.94 (7.96)30.85 (7.74)28.92 (7.60)29.36 (6.91)28.66 (6.91)28.97 (7.29)
Ct, median (IQR)24.32 (20.60–34.68)33.56 (23.31–37.80)28.19 (22.18–36.20)28.57 (23.51–36.09)27.93 (22.74–35.02)28.36 (22.60–35.87)
Ct ≤24, %46.828.034.227.431.232.0

Abbreviations: Ct, cycle threshold; IQR, interquartile range (25th–75th percentiles); SD, standard deviation. † Recorded age. Within ages 1–17 years recorded age and first test date are jointly ordered (Table 2), so these columns cannot be read as true age groups. Note: Recorded collection day is the day of the time-shifted 2020 pandemic period (4–107). Lower Ct means more viral RNA.

Recorded age was ordered by first test day in every band (Table 2): ρ was 0.801 among infants, 0.975 at ages 1–17 years and 0.644 among adults. Within-band age is therefore unusable in all 3 bands. Band membership agreed with the 2 available checks (97.6% of 294 NICU tests recorded as <1 year; all 12 employee-health tests as ≥18 years), though NICU recorded ages spread across the first year. Recorded sex was altered (56.5% of OB-GYN tests recorded as female).

The repeat-tester check is clear on 2 points and silent on a third:

  • Recorded age never changed within any of 1,744 repeat testers, including 192 infants whose tests spanned at least 37 days. If age were recorded at each collection, 254.1, 103.6 and 49.1 changes would be expected in infants, children and adults. Recorded age is a per-person value.
  • Within-person gaps had a median of 16.0 days (IQR 6.0–33.0), and only 7.3% were 1 day or less. Dates were therefore not re-sorted test by test.
  • Recorded age tracked each person's first test day (ρ 0.903 among children) much more than the last (0.276). This fits ages reassigned in first-test order, but whole per-person date sets re-placed by age would produce the same pattern, so the check cannot separate the 2 readings.

Daily counts had a strong weekly cycle (lag-7 autocorrelation 0.679), yet only 1 of 33 clinics had a weekday profile departing from the overall one, about what chance would give (1.65 expected). Clinics that ordinarily operate on weekdays only had tests on every recorded weekday (Table 2). Ct retained an association with setting (R² 0.0502, against a permutation 95th percentile of 0.0122) but showed none beyond chance with recorded date (0.0080, against 0.0097). That check has limited power against a modest epidemic-phase trend (3-df spline, 785 specimens), so the conclusion that recorded dates are detached from the records' clinical content rests mainly on the weekday-only clinics. Whichever field was reordered, the spline in M2 and M3 cannot be read as calendar time, and calendar time cannot be adjusted for.

Table 2. What Survived Deidentification: Checks of Recorded Age, Recorded Date and Ct, Children's Hospital of Philadelphia Testing Records, 2020

CheckResult
Spearman ρ, recorded age vs first test day: infants (1,500 persons); ages 1–17 y (7,179); adults (3,665)0.801; 0.975; 0.644
Median recorded age by first test on days 4–37 / 38–72 / 73–107: infants; 1–17 y; adults0.1 / 0.4 / 0.8; 2.0 / 6.0 / 14.0; 25.0 / 33.0 / 47.0 y
R² of recorded age on recorded-day spline, ages 1–17 y (analytic sample)0.671
NICU tests recorded <1 y; employee-health tests recorded ≥18 y97.6% of 294; 100.0% of 12
NICU recorded age, 10th / 50th / 90th percentile0.1 / 0.5 / 0.8 y
OB-GYN tests recorded female56.5% of 69
Repeat testers: persons; tests; recorded age constant within person1,744; 4,924; 100.0%
Infants whose tests span ≥37 days; of these, with any change in recorded age192; 0
Changes expected if age were recorded at each collection: infants; children; adults (observed: 0; 0; 0)254.1; 103.6; 49.1
Days between consecutive tests: median (IQR); share ≤1 day; share ≥28 days16.0 (6.0–33.0); 7.3%; 31.1%
Repeat testers, ρ of recorded age with first vs last test day: infants; children; adults0.431 vs 0.458; 0.903 vs 0.276; 0.551 vs 0.149
Detrended daily test counts, lag-7 autocorrelation0.679
Clinics (≥50 tests) whose weekday profile departs from the overall profile (χ² > 12.59, 6 df)1 of 33 (1.65 expected by chance)
Two lowest weekday shares: radiation oncology (172 tests); oncology day hospital (533); hematology-oncology day hospital (71)9.9%, 12.2%; 12.0%, 12.8%; 2.8%, 9.9%
R² of Ct on recorded-day spline only: observed vs permutation 95th / 99th percentile0.0080 vs 0.0097 / 0.0132
R² of Ct on setting only: observed vs permutation 95th / 99th percentile0.0502 vs 0.0122 / 0.0172
R² of Ct on setting and recorded-day spline: observed vs permutation 95th / 99th percentile0.055 vs 0.018 / 0.022

Abbreviations: Ct, cycle threshold; df, degrees of freedom; NICU, neonatal intensive care unit; OB-GYN, obstetrics and gynecology; R², coefficient of determination. Note: The age correlations, medians and checks against clinical structure use 1 row per person or all 15,524 tests. Repeat testers are persons with 2 or more tests of any result. Ct checks use the 785 analytic specimens. The expected changes count infants whose tests span ≥37 days plus fractional spans, and birthdays within each person's span.

Without recorded dates, infants' Ct was 1.72 cycles lower than adults' unadjusted (95% CI: −3.65, 0.20) and 2.11 cycles lower in the setting-only model (95% CI: −4.04, −0.18) (Table 3). That corresponds to 4.32-fold more RNA (95% CI: 1.13, 16.45). The pre-specified M3 gave a similar −2.25 (95% CI: −4.20, −0.31), as expected if the spline carries little information about Ct. For persons aged 1–17 years the setting-only difference was 0.73 (95% CI: −0.39, 1.85), compatible with Ct values from 0.39 cycles lower to 1.85 cycles higher than adults'. The recorded 1–4-year band (unadjusted 2.19; 95% CI: 0.53, 3.85) and the composite under-5 group (M3, −0.52; 95% CI: −2.06, 1.03) are not age contrasts.

Table 3. Mean Difference in Cycle Threshold (Cycles) Relative to Adults, by Age Coding and Model, Children's Hospital of Philadelphia, 2020

Age group (coding)No.M1, unadjustedSetting onlyM2, recorded-day splineM3, spline + setting
Infants, <1 y (band)79−1.72 (−3.65, 0.20)−2.11 (−4.04, −0.18)−1.94 (−3.88, 0.00)−2.25 (−4.20, −0.31)
Ages 1–17 y (band)3920.97 (−0.09, 2.03)0.73 (−0.39, 1.85)0.93 (−0.13, 1.99)0.72 (−0.40, 1.85)
Under 5 y (composite)1860.53 (−0.86, 1.92)—−0.13 (−1.60, 1.34)−0.52 (−2.06, 1.03)
5–17 y (composite)2850.52 (−0.62, 1.65)—0.80 (−0.35, 1.95)0.74 (−0.46, 1.94)
Infants (5-band model)79−1.72 (−3.65, 0.20)—−1.92 (−3.87, 0.02)−2.20 (−4.16, −0.24)
Recorded 1–4 y (5-band model)1072.19 (0.53, 3.85)—1.53 (−0.31, 3.36)1.13 (−0.78, 3.05)
Recorded 5–11 y (5-band model)1170.26 (−1.32, 1.84)—0.36 (−1.34, 2.05)0.75 (−0.99, 2.49)
Recorded 12–17 y (5-band model)1680.70 (−0.60, 1.99)—0.94 (−0.41, 2.29)0.51 (−0.89, 1.91)

Note: Values are mean differences (95% confidence interval) from ordinary least squares with HC3 standard errors. The reference is adults (n = 314; mean, 28.66). Negative values mean lower cycle thresholds (more RNA). Only the 2 band rows are age contrasts. The recorded-day spline in M2 and M3 cannot be read as calendar time. The setting-only model, added in revision, was fitted to the band coding only. M3 was the pre-specified primary model.

The PR for Ct ≤24 was 1.50 (95% CI: 1.13, 2.00) unadjusted and 1.51 (95% CI: 1.11, 2.06) with setting adjustment (Table 4). Without the spline, the infant–adult difference was −2.64 (95% CI: −5.71, 0.43) within the drive-through and −1.59 (95% CI: −4.36, 1.19) elsewhere, and sensitivity estimates ranged from −1.77 to −2.90. With all 856 specimens it was −1.77 (95% CI: −3.60, 0.06), compatible with a deficit of up to 3.60 cycles or none. M3 versions were similar. The date-free adjusted estimate is very fragile. A confounder explaining 0.97% of the residual variance of both infant status and Ct would move its conventional-standard-error interval to include 0. In the same date-free model, setting alone has partial R² values of 5.11% with Ct and 10.45% with infant status. Infant selection on viral load at an OR of 1.20 per 5 cycles would remove the difference, and selective testing of adults would enlarge it.

Table 4. Secondary, Sensitivity and Bias Analyses for Infants Versus Adults Under the Setting-Only and M3 Specifications, Children's Hospital of Philadelphia, 2020

AnalysisSampleSetting only (95% CI)M3 (95% CI)
Fold difference in RNA (2−Δ2^{-\Delta})785 persons4.32 (1.13, 16.45)4.76 (1.24, 18.38)
PR of Ct ≤24 (unadjusted M1: 1.50; 1.13, 2.00)785 persons1.51 (1.11, 2.06)1.51 (1.11, 2.06)
E-value for the PR: estimate; confidence limit—2.39; 1.462.39; 1.46
PR of Ct ≤24, under 5 y (composite) vs adults: not interpretable as an age contrast785 persons—1.27 (0.95, 1.70)
Drive-through only22 infants, 251 adults−2.64 (−5.71, 0.43)−2.73 (−5.77, 0.30)
Other settings57 infants, 63 adults−1.59 (−4.36, 1.19)−1.90 (−4.79, 0.99)
Infant × drive-through product term785 persons−1.05 (−5.19, 3.09)−1.08 (−5.22, 3.07)
Weekly indicators instead of spline: check that the time coding is immaterial785 persons—−2.25 (−4.20, −0.30)
All quantifiable positive specimens, person-clustered SEs856 specimens, 785 persons−1.77 (−3.60, 0.06)−1.79 (−3.63, 0.06)
Excluding outside institutions711 persons: 69 infants, 286 adults−2.46 (−4.54, −0.38)−2.56 (−4.66, −0.45)
Hospital patients only484 persons, 76 adults−2.90 (−5.36, −0.44)−2.83 (−5.32, −0.33)
Robustness value: estimate to 0; interval to include 0785 persons7.71%; 0.97%8.18%; 1.47%
Partial R², same model: infant status with Ct; setting with Ct; setting with infant status785 persons0.64%; 5.11%; 10.45%0.72%; 4.88%; 10.22%
Selection OR per 5 cycles lower Ct: infants 1.25 (adults 1)785 persons0.50 (−1.54, 2.55)0.32 (−1.72, 2.37)
Selection OR: infants 1.5 (adults 1)785 persons2.54 (0.52, 4.57)2.33 (0.31, 4.34)
Selection OR: infants 2 (adults 1)785 persons5.11 (3.27, 6.96)4.84 (3.00, 6.68)
Selection OR: adults 1.5 (infants 1)785 persons−5.66 (−7.55, −3.76)−5.74 (−7.66, −3.83)
Selection OR: adults 2 (infants 1)785 persons−7.58 (−9.45, −5.70)−7.63 (−9.52, −5.75)
Infant selection OR at which the difference reaches 0785 persons1.201.22

Abbreviations: CI, confidence interval; Ct, cycle threshold; OR, odds ratio; PR, prevalence ratio; R², coefficient of determination; SE, standard error. Note: Unless stated otherwise, entries are mean differences in cycles, infants minus adults. The setting-only column omits the recorded-day spline, and the M3 column keeps it. Within the drive-through, setting is constant, so the setting-only estimate there is unadjusted and the M3 estimate is spline adjusted. PRs come from modified Poisson regression with robust variance, and the fold difference assumes 100% amplification efficiency. Robustness values follow Cinelli and Hazlett (29) and use the conventional ordinary least squares t statistic, so their interval criterion refers to a conventional-standard-error interval rather than the HC3 interval. Selection rows give bias-corrected differences under the exponential-tilt model.

Discussion

These records cannot answer the question as posed. Recorded age is ordered by date within every band, and recorded dates appear detached from the records' clinical content, so neither children aged 1–4 years nor calendar time can be studied. The checks favor reassignment of ages over test-by-test re-sorting of dates, but whole per-person date sets re-placed by age remain possible. Under either reading, the calendar-time-adjusted estimate (M3, −2.25) is not interpretable as an adjustment for calendar time. Under the date-reordering reading, the relevant unadjusted estimate is −1.72 (95% CI: −3.65, 0.20), whose interval includes 0. Adjustment for setting, the only adjustment the records support, gives −2.11 (95% CI: −4.04, −0.18), close to 0 at its upper limit and very fragile. The data are therefore compatible with a small infant excess, of roughly 2 cycles, but do not establish it. For persons aged 1–17 years as a group, differences were small (0.73; 95% CI: −0.39, 1.85).

The direction of the infant estimates agrees with studies that separated infants (10, 11). Their size is far below the 10- to 100-fold under-5 excess reported from Chicago, which was restricted to symptomatic patients within a week of onset (2). The estimate for persons aged 1–17 years fits large series in which children's loads were similar to or lower than adults' (3–6, 8), and the small difference between symptomatic children and adults in community surveillance (9). Setting mattered here as elsewhere (12): Ct was associated with setting, and setting adjustment moved the infant estimate from −1.72 to −2.11.

The strengths of the study are explicit checks of what deidentification preserved, and quantification of both unmeasured confounding and selection.

Confounding. The records lack time since symptom onset, symptom status and testing indication, the strongest known determinants of Ct (9, 12), and setting is only a coarse proxy. A confounder much weaker than setting could move the date-free conventional-standard-error interval to include 0. Because setting partly results from age and illness, the setting-adjusted contrast is a conditional description, not an effect of age.

Selection. Everyone analyzed was tested and tested positive (16, 17, 30, 31). If infected infants were tested mainly when ill, while adults with low loads were also found by screening, the infant excess is overstated, and a relative selection OR of 1.20 per 5 cycles would remove it. If adults were tested mainly when symptomatic, the excess is understated. The adult group is unusual as well: 79.9% were tested at the drive-through, and restricting to hospital patients left only 76 adults.

Information bias. The deidentification is the largest problem, and it is not classical nondifferential misclassification. Within every band, recorded age and date are jointly ordered, and both appear detached from the records' clinical content. Their effect on within-band contrasts is unpredictable, as warned for swapping-based disclosure control (19). The weekday reasoning rests mainly on clinics that ordinarily operate on weekdays only and assumes they did so in 2020; the Ct–date check has limited power. The checks support band membership and rule out a full permutation of Ct, but permutation of Ct within bands or strata cannot be tested; all estimates assume it did not occur. Ct is assay- and specimen-specific (23), the platform mix may differ by age, and smaller swabs for the youngest children lowered measured loads in one large series (4). The Ct ≤24 threshold comes from an E-gene assay (24), and the fold conversion assumes perfect efficiency.

Precision and analytic choices. Only 79 infants were analyzed. The question about children younger than 5 years was narrowed to infants by a rule fixed after inspecting the data, and the pre-specified criterion, an interval excluding 0, was met only by thin margins. The setting-only model, its robustness analyses and the date checks were added at review. The data come from one center in 2020.

For public health, these records support at most a small excess of upper-respiratory RNA in infants. They cannot inform policies aimed at toddlers or at children younger than 5 years as a whole. More generally, teaching datasets can carry undocumented perturbations that manufacture or erase gradients by age and date. Repeat-tester, weekday and permutation checks should precede any measure of association, and data donors should document which fields were permuted.

Acknowledgments

Funding: this work received no specific funding. Conflict of interest: none declared. Ethics and data use: this is a secondary analysis of publicly distributed, deidentified records. The dataset is redistributed in the R package medicaldata under the package's MIT license, with no separate data license stated. No persons were contacted, and no ethics review was sought. Data and code availability: the data are available at https://vincentarelbundock.github.io/Rdatasets/csv/medicaldata/covid_testing.csv (22). The executed analysis script and its complete output are archived with this study's run record (identifier 10fc99e05068), together with the data file's SHA-256 checksum (fc7d7c522749144b…); no digital object identifier has been issued.

Abbreviations

CHOP, Children's Hospital of Philadelphia; CI, confidence interval; Ct, cycle threshold; df, degrees of freedom; HC3, heteroskedasticity-consistent covariance estimator 3; IQR, interquartile range; NICU, neonatal intensive care unit; OB-GYN, obstetrics and gynecology; OR, odds ratio; PR, prevalence ratio; R², coefficient of determination; RECORD, REporting of studies Conducted using Observational Routinely-collected health Data; RT-PCR, reverse-transcription polymerase chain reaction; SARS-CoV-2, severe acute respiratory syndrome coronavirus 2; SD, standard deviation; SE, standard error; STROBE, Strengthening the Reporting of Observational Studies in Epidemiology.

References

  1. Paul LA, Daneman N, Schwartz KL, et al. Association of age and pediatric household transmission of SARS-CoV-2 infection. JAMA Pediatr. 2021;175(11):1151–1158. https://doi.org/10.1001/jamapediatrics.2021.2770
  2. Heald-Sargent T, Muller WJ, Zheng X, et al. Age-related differences in nasopharyngeal severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) levels in patients with mild to moderate coronavirus disease 2019 (COVID-19). JAMA Pediatr. 2020;174(9):902–903. https://doi.org/10.1001/jamapediatrics.2020.3651
  3. Euser S, Aronson S, Manders I, et al. SARS-CoV-2 viral-load distribution reveals that viral loads increase with age: a retrospective cross-sectional cohort study. Int J Epidemiol. 2021;50(6):1795–1803. https://doi.org/10.1093/ije/dyab145
  4. Jones TC, Biele G, Mühlemann B, et al. Estimating infectiousness throughout SARS-CoV-2 infection course. Science. 2021;373(6551):eabi5273. https://doi.org/10.1126/science.abi5273
  5. Madera S, Crawford E, Langelier C, et al. Nasopharyngeal SARS-CoV-2 viral loads in young children do not differ significantly from those in older children and adults. Sci Rep. 2021;11(1):3044. https://doi.org/10.1038/s41598-021-81934-w
  6. Baggio S, L'Huillier AG, Yerly S, et al. Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) viral load in the upper respiratory tract of children and adults with early acute coronavirus disease 2019 (COVID-19). Clin Infect Dis. 2021;73(1):148–150. https://doi.org/10.1093/cid/ciaa1157
  7. Yonker LM, Neilan AM, Bartsch Y, et al. Pediatric severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2): clinical presentation, infectivity, and immune responses. J Pediatr. 2020;227:45–52.e5. https://doi.org/10.1016/j.jpeds.2020.08.037
  8. Bullard J, Funk D, Dust K, et al. Infectivity of severe acute respiratory syndrome coronavirus 2 in children compared with adults. CMAJ. 2021;193(17):E601–E606. https://doi.org/10.1503/cmaj.210263
  9. Chung E, Chow EJ, Wilcox NC, et al. Comparison of symptoms and RNA levels in children and adults with SARS-CoV-2 infection in the community setting. JAMA Pediatr. 2021;175(10):e212025. https://doi.org/10.1001/jamapediatrics.2021.2025
  10. Ochoa V, Díaz FE, Ramirez E, et al. Infants younger than 6 months infected with SARS-CoV-2 show the highest respiratory viral loads. J Infect Dis. 2022;225(3):392–395. https://doi.org/10.1093/infdis/jiab577
  11. Pinninti SG, Pati S, Poole C, et al. Virological characteristics of hospitalized children with SARS-CoV-2 infection. Pediatrics. 2021;147(5):e2020037812. https://doi.org/10.1542/peds.2020-037812
  12. Kociolek LK, Muller WJ, Yee R, et al. Comparison of upper respiratory viral load distributions in asymptomatic and symptomatic children diagnosed with SARS-CoV-2 infection in pediatric hospital testing programs. J Clin Microbiol. 2020;59(1):e02593-20. https://doi.org/10.1128/JCM.02593-20
  13. Otto WR, Geoghegan S, Posch LC, et al. The epidemiology of severe acute respiratory syndrome coronavirus 2 in a pediatric healthcare network in the United States. J Pediatric Infect Dis Soc. 2020;9(5):523–529. https://doi.org/10.1093/jpids/piaa074
  14. Hay JA, Kennedy-Shaffer L, Kanjilal S, et al. Estimating epidemiologic dynamics from cross-sectional viral load distributions. Science. 2021;373(6552):eabh0635. https://doi.org/10.1126/science.abh0635
  15. Walker AS, Pritchard E, House T, et al. Ct threshold values, a proxy for viral load in community SARS-CoV-2 cases, demonstrate wide variation across populations and over time. eLife. 2021;10:e64683. https://doi.org/10.7554/eLife.64683
  16. Hernán MA, Hernández-Díaz S, Robins JM. A structural approach to selection bias. Epidemiology. 2004;15(5):615–625. https://doi.org/10.1097/01.ede.0000135174.63482.43
  17. Griffith GJ, Morris TT, Tudball MJ, et al. Collider bias undermines our understanding of COVID-19 disease risk and severity. Nat Commun. 2020;11(1):5749. https://doi.org/10.1038/s41467-020-19478-2
  18. Higgins PDR. covid_testing: COVID-19 testing data from CHOP, 2020. medicaldata package reference documentation. https://higgi13425.github.io/medicaldata/reference/covid_testing.html. Accessed October 24, 2026.
  19. Dalenius T, Reiss SP. Data-swapping: a technique for disclosure control. J Stat Plan Inference. 1982;6(1):73–85. https://doi.org/10.1016/0378-3758(82)90058-1
  20. von Elm E, Altman DG, Egger M, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Epidemiology. 2007;18(6):800–804. https://doi.org/10.1097/EDE.0b013e3181577654
  21. Benchimol EI, Smeeth L, Guttmann A, et al. The REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) statement. PLoS Med. 2015;12(10):e1001885. https://doi.org/10.1371/journal.pmed.1001885
  22. Higgins PDR. medicaldata: Data Package for Medical Datasets. R package version 0.2.0. 2021. https://CRAN.R-project.org/package=medicaldata. Data file covid_testing (donated by Amrom E. Obstfeld), Rdatasets archive: https://vincentarelbundock.github.io/Rdatasets/csv/medicaldata/covid_testing.csv. Accessed October 24, 2026.
  23. Rhoads D, Peaper DR, She RC, et al. College of American Pathologists (CAP) Microbiology Committee perspective: caution must be used in interpreting the cycle threshold (Ct) value. Clin Infect Dis. 2021;72(10):e685–e686. https://doi.org/10.1093/cid/ciaa1199
  24. Bullard J, Dust K, Funk D, et al. Predicting infectious severe acute respiratory syndrome coronavirus 2 from diagnostic samples. Clin Infect Dis. 2020;71(10):2663–2666. https://doi.org/10.1093/cid/ciaa638
  25. MacKinnon JG, White H. Some heteroskedasticity-consistent covariance matrix estimators with improved finite sample properties. J Econom. 1985;29(3):305–325. https://doi.org/10.1016/0304-4076(85)90158-7
  26. Westreich D, Greenland S. The Table 2 fallacy: presenting and interpreting confounder and modifier coefficients. Am J Epidemiol. 2013;177(4):292–298. https://doi.org/10.1093/aje/kws412
  27. Zou G. A modified Poisson regression approach to prospective studies with binary data. Am J Epidemiol. 2004;159(7):702–706. https://doi.org/10.1093/aje/kwh090
  28. VanderWeele TJ, Ding P. Sensitivity analysis in observational research: introducing the E-value. Ann Intern Med. 2017;167(4):268–274. https://doi.org/10.7326/M16-2607
  29. Cinelli C, Hazlett C. Making sense of sensitivity: extending omitted variable bias. J R Stat Soc Series B Stat Methodol. 2020;82(1):39–67. https://doi.org/10.1111/rssb.12348
  30. Accorsi EK, Qiu X, Rumpler E, et al. How to detect and reduce potential sources of biases in studies of SARS-CoV-2 and COVID-19. Eur J Epidemiol. 2021;36(2):179–196. https://doi.org/10.1007/s10654-021-00727-7
  31. Smith LH, VanderWeele TJ. Bounding bias due to selection. Epidemiology. 2019;30(4):509–516. https://doi.org/10.1097/EDE.0000000000001032
  32. Scharfstein DO, Rotnitzky A, Robins JM. Adjusting for nonignorable drop-out using semiparametric nonresponse models. J Am Stat Assoc. 1999;94(448):1096–1120. https://doi.org/10.1080/01621459.1999.10473862
  33. Lash TL, Fox MP, MacLehose RF, et al. Good practices for quantitative bias analysis. Int J Epidemiol. 2014;43(6):1969–1985. https://doi.org/10.1093/ije/dyu149

6.Review

Attempt 2Done

Automated review: Accept (novelty 3, rigor 4, clarity 4, reproducibility 4).

Accept
Novelty 3/5Rigor 4/5Clarity 4/5Reproducibility 4/5

Refereed by Claude Fable 5.1, a different model from the one that wrote the paper (Claude Opus 5.5).

Across three rounds the author has met every required change with analyses that were actually run and reported exactly as executed. The round-3 re-run aligns the stratified, product-term, sensitivity and selection analyses and their benchmarks with the date-free setting-only model that the paper now reports, and the results change little, as expected given that the recorded-date spline explains almost nothing of Ct. The weekday evidence is weighted as it should be, the ED-vs-drive-through comparison is removed, the power limit of the Ct-date check is stated, and the Yonker citation is corrected. The manuscript agrees with the experiment record in every number and design choice I checked, follows the journal's conventions, and claims no more than the data support: a small infant excess in upper-respiratory RNA that the records are compatible with but do not establish, and a clear demonstration that this teaching dataset cannot answer the under-5 question once calendar time is considered. The remaining notes (code deposit with a persistent identifier, a drawn flow diagram) are for production. I recommend acceptance.

Strengths

  • Round-3 verification against the re-executed record (script now 291 lines; data file and loader unchanged by checksum). The three changes I required in round 2 are all met. (1) The robustness analyses were re-run without recorded-date terms (script section 11): within drive-through -2.64 (95% CI: -5.71, 0.43), other settings -1.59 (-4.36, 1.19), product term -1.05 (-5.19, 3.09), all 856 specimens with person-clustered SEs -1.77 (-3.60, 0.06), excluding outside institutions -2.46 (-4.54, -0.38), hospital patients only -2.90 (-5.36, -0.44), selection scenarios 0.50 to 5.11 (infant ORs) and -5.66 to -7.58 (adult ORs), tipping point 1.20, and date-free partial R-squared benchmarks of 5.11% (setting with Ct) and 10.45% (setting with infant status) against the 0.97% robustness value. Every entry in Table 4's setting-only column and in the Results text matches RESULT_JSON, and the Methods, Table 4 note (setting constant within the drive-through, so that stratum is unadjusted) and Discussion describe the specification exactly as executed. (2) The weekday evidence is stated as 1 of 33 clinics with 1.65 expected by chance (0.05 x 33, now computed by the script), the emergency-department vs drive-through comparison is gone from the text, Methods and Table 2, and the limited power of the Ct-date check (3-df spline, 785 specimens) is stated in Results and in the information-bias paragraph. (3) Yonker is cited for no age gradient within the cohort with the note that early loads exceeded hospitalized adults', and is removed from the 'similar to or lower than adults' group (3-6, 8).
  • The manuscript now agrees with the record in every number I checked across all three rounds: sample flow and missing values by stage, Table 1, all four age-coding models, the repeat-tester, weekday and split permutation checks, the date-free estimates, and both columns of the robustness table. Derived figures recompute (2^2.11 = 4.32 with limits 1.13 and 16.45; E-values 2.39 and 1.46; 0.05 x 33 = 1.65; sensitivity range -1.77 to -2.90).
  • The inferential position is honest and proportionate: the records cannot answer the under-5 question; calendar time cannot be adjusted for under either reading of the perturbation; the only supportable adjustment (setting) gives -2.11 (95% CI: -4.04, -0.18), described as very fragile, with a tipping-point selection OR of 1.20; the conclusion is 'compatible with a small infant excess' and no more. The pre-specified M3 result is reported but not relied on, and everything added at review is labelled as such.
  • The diagnostic sequence for perturbed open datasets (age-date correlation, clinical-structure checks, repeat-tester constancy and gaps, weekday profiles of weekday-only clinics, split permutation nulls for the outcome) is the paper's durable contribution and is now fully documented and reproducible from the archived script and output.
  • Venue conventions are met: measures of association with 95% CIs throughout, no 'significant' or p-value-centred language, abbreviations defined at first use, STROBE/RECORD flow with exclusions and missing data, funding, conflict-of-interest and ethics statements, software versions and seed; abstract exactly 200 words; all 33 references are among the retrieved sources and resolve.

Weaknesses

  • Remaining limitations are inherent to the data and are acknowledged: 79 infants; an unusual adult comparison group (79.9% drive-through, 76 hospital patients); a decision rule set after inspecting the data; permutation of Ct within bands untestable; the Ct <=24 marker transferred from an E-gene assay; one center in 2020. The substantive result is correspondingly modest.
  • Code availability points to the run-record identifier and data checksum because no DOI exists; the script and output should be deposited with a persistent identifier at production. Figure 1 is a table-style flow because the record renders only Markdown tables; production will need a drawn flow diagram.
  • The 38.2% vs 7.2% asymptomatic proportions attributed to ref 9 (Chung et al.) could not be verified from the retrieved record within this review; the author reports having checked them against the paper.
  • Table 2 is dense (18 rows of heterogeneous checks); a two-part layout (age and date checks; Ct checks) would help readers, but this is presentational.

Required changes

—

Record

§Event timeline

  1. Run createdQuestion: “Do children under 5 years have higher SARS-CoV-2 upper-respiratory viral loads (lower RT-PCR cycle threshold values) than adults, once testing setting and calendar time are accounted for?' language=en discipline=epidemiology venue='American Journal of Epidemiology” · output language: English · models: agent:claude-opus-5-5
  2. Stage started · Literature reviewAttempt 1 · agent:claude-opus-5-5
  3. awaiting_agent · Literature reviewsearch response for literature_review: literature_review-search-1.response.json
  4. Resume requested
  5. Stage started · Literature reviewAttempt 2 · agent:claude-opus-5-5
  6. Search complete · Literature review46 sources · 18,395 characters of notes
  7. awaiting_agent · Literature reviewcomplete_json response for literature_review: literature_review-complete_json-1.response.json
  8. Resume requested
  9. Stage started · Literature reviewAttempt 3 · agent:claude-opus-5-5
  10. Search complete · Literature review46 sources · 18,395 characters of notes
  11. Stage completed · Literature review
  12. Stage started · IdeationAttempt 1 · agent:claude-opus-5-5
  13. awaiting_agent · Ideationcomplete_json response for ideation: ideation-complete_json-1.response.json
  14. Resume requested
  15. Stage started · IdeationAttempt 2 · agent:claude-opus-5-5
  16. Awaiting approval · Ideation
  17. Approved · IdeationIdea 1 selected
  18. Stage started · ExperimentationAttempt 1 · agent:claude-opus-5-5
  19. Dataset provisioned · Experimentationchop_covid_testing · 15,524 observations · 15 variables · sha256 fc7d7c522749
  20. awaiting_agent · Experimentationcomplete response for experimentation: experimentation-complete-1.response.txt
  21. Resume requested
  22. Stage started · ExperimentationAttempt 2 · agent:claude-opus-5-5
  23. Dataset provisioned · Experimentationchop_covid_testing · 15,524 observations · 15 variables · sha256 fc7d7c522749
  24. Experiment round · ExperimentationRound 1: exit 0; result parsed (1.218s)
  25. Stage completed · Experimentation
  26. Stage started · AnalysisAttempt 1 · agent:claude-opus-5-5
  27. awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
  28. Resume requested
  29. Stage started · AnalysisAttempt 2 · agent:claude-opus-5-5
  30. Stage completed · Analysis
  31. Stage started · Paper writingAttempt 1 · agent:claude-opus-5-5
  32. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  33. Resume requested
  34. Stage started · Paper writingAttempt 2 · agent:claude-opus-5-5
  35. Stage completed · Paper writing
  36. Stage started · ReviewAttempt 1 · agent:claude-opus-5-5
  37. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  38. Resume requested
  39. Stage started · ReviewAttempt 2 · agent:claude-opus-5-5
  40. Awaiting approval · Review
  41. Sent back for revision · ReviewRe-running from Paper writing: “Referee recommendation: minor_revision. Make every required change below, using only results the executed script already produced (no new analyses, no new numbers):\n1. "Replace the abstract\'s final cl”
  42. Stage started · Paper writingAttempt 3 · agent:claude-opus-5-5
  43. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  44. Resume requested
  45. Stage started · Paper writingAttempt 4 · agent:claude-opus-5-5
  46. Stage completed · Paper writing
  47. Stage started · ReviewAttempt 1 · agent:claude-opus-5-5
  48. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  49. Resume requested
  50. Stage started · ReviewAttempt 2 · agent:claude-opus-5-5
  51. Awaiting approval · Review
  52. Approved · Review
  53. Run completereport=/Users/haodingresearch/Desktop/auto-research/runs/studies-opus55/epidemiology-en/runs/10fc99e05068/REPORT.md
  54. Reopened after sign-off · ReviewReopened after sign-off; re-running from Review: “Independent review: the paper is re-refereed by an independent referee (Claude Fable 5.1) that took no part in writing it.”
  55. Stage started · ReviewAttempt 3 · agent:claude-fable-5-1
  56. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  57. Resume requested
  58. Stage started · ReviewAttempt 4 · agent:claude-fable-5-1
  59. Awaiting approval · Review
  60. Sent back for revision · ReviewRe-running from Experimentation: “Independent referee (Claude Fable 5.1): major_revision. Required changes: (1) [analysis] Run the discriminating check you name: among persons with 2 or more tests (any result), report whether recorded”
  61. Stage started · ExperimentationAttempt 3 · agent:claude-opus-5-5
  62. Dataset provisioned · Experimentationchop_covid_testing · 15,524 observations · 15 variables · sha256 fc7d7c522749
  63. awaiting_agent · Experimentationcomplete response for experimentation: experimentation-complete-1.response.txt
  64. Resume requested
  65. Stage started · ExperimentationAttempt 4 · agent:claude-opus-5-5
  66. Dataset provisioned · Experimentationchop_covid_testing · 15,524 observations · 15 variables · sha256 fc7d7c522749
  67. Experiment round · ExperimentationRound 1: exit 0; result parsed (1.429s)
  68. Stage completed · Experimentation
  69. Stage started · AnalysisAttempt 1 · agent:claude-opus-5-5
  70. awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
  71. Resume requested
  72. Stage started · AnalysisAttempt 2 · agent:claude-opus-5-5
  73. Stage completed · Analysis
  74. Stage started · Paper writingAttempt 5 · agent:claude-opus-5-5
  75. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  76. Resume requested
  77. Stage started · Paper writingAttempt 6 · agent:claude-opus-5-5
  78. Stage completed · Paper writing
  79. Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
  80. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  81. Resume requested
  82. Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
  83. Awaiting approval · Review
  84. Sent back for revision · ReviewRe-running from Experimentation: “Independent referee (Claude Fable 5.1), round 2: minor_revision. Required changes: (1) [analysis] Re-run the within-setting contrasts, the infant x drive-through product term, the three applicable sen”
  85. Stage started · ExperimentationAttempt 5 · agent:claude-opus-5-5
  86. Dataset provisioned · Experimentationchop_covid_testing · 15,524 observations · 15 variables · sha256 fc7d7c522749
  87. awaiting_agent · Experimentationcomplete response for experimentation: experimentation-complete-1.response.txt
  88. Resume requested
  89. Stage started · ExperimentationAttempt 6 · agent:claude-opus-5-5
  90. Dataset provisioned · Experimentationchop_covid_testing · 15,524 observations · 15 variables · sha256 fc7d7c522749
  91. Experiment round · ExperimentationRound 1: exit 0; result parsed (1.465s)
  92. Stage completed · Experimentation
  93. Stage started · AnalysisAttempt 1 · agent:claude-opus-5-5
  94. awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
  95. Resume requested
  96. Stage started · AnalysisAttempt 2 · agent:claude-opus-5-5
  97. Stage completed · Analysis
  98. Stage started · Paper writingAttempt 7 · agent:claude-opus-5-5
  99. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  100. Resume requested
  101. Stage started · Paper writingAttempt 8 · agent:claude-opus-5-5
  102. Stage completed · Paper writing
  103. Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
  104. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  105. Resume requested
  106. Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
  107. Awaiting approval · Review
  108. Approved · Review
  109. Run completereport=/Users/haodingresearch/Desktop/auto-research/runs/studies-opus55/epidemiology-en/runs/10fc99e05068/REPORT.md