An independent audit of published HFMD surveillance statistics and a documented birth-date generation mechanism in China’s disease-reporting system.
The Null Institute | Research Note & Verification Audit | 6 November 2026 | Version 1.0.
Scientific status: Public working research note, not peer-reviewed. No individual Yunnan patient records were accessed. The source-study authors have not been contacted.
Key result: We identified a reproducible discrepancy in a published chi-square statistic, two contradictions between reported crude proportions and the stated main-effects regression, and calendar-code defects. Independently, historical Chinese surveillance sources confirm that software-generated dates of birth occurred in real reporting records. Whether that mechanism created the reported sixfold HFMD clustering in Yunnan remains unproven.
Research files & reproducibility
Public release v1.0, 6 November 2026. Independently checked downloads with SHA-256.
The CSV contains numbers already published by Jiang et al. (2024), not original medical records. The ZIP contains Institute-authored checks only, not the authors’ R script or patient data. This note has not been peer-reviewed.
The reported birthday-week anomaly
Jiang et al. (2024) counted 973,410 surveillance cases, including 116,976 with recorded symptom onset on the birthday or in the preceding six days (12.02%). The seven individual counts from day −6 through day 0 are 3,308, 4,442, 8,028, 14,982, 25,945, 41,930 and 18,341. The concentration is extremely nonuniform, with its maximum one day before the recorded birthday. The percentage of cases in the window declined from 33.74% in 2008 (3,914/11,600) to 2.77% in 2022 (2,254/81,372). All figures in this paragraph are those of the original paper, not estimates from independently obtained Yunnan records.
A numerator of infections among registered HFMD cases in a window, divided by all registered HFMD cases, is a case-only timing statistic. Comparing it with 7/365 does not, by itself, estimate the causal risk of disease around independently verified birthdays. Potential biases include nonuniform birthdates, strong age dependence of HFMD, seasonal incidence, ascertainment, date imputation and differences between symptom onset and medical reporting. It would nevertheless be incorrect to assume these factors explain the sixfold concentration without testing them.
| Day relative to recorded birthday | Published cases |
|---|---|
| −6 | 3,308 |
| −5 | 4,442 |
| −4 | 8,028 |
| −3 | 14,982 |
| −2 | 25,945 |
| −1 | 41,930 |
| 0 (birthday) | 18,341 |
Published case counts from Jiang et al. (2024), Figure 1B. The rise towards day −1 is present in the authors’ data; its biological or administrative cause has not been established.
1. Checks that can be reproduced without patient records
Pearson χ² for sex. The paper’s Table 1 reports 68,849 birthday-week cases among 564,725 boys and 48,127 among 408,685 girls. A conventional uncorrected Pearson 2×2 contingency test yields χ²(1) = 38.7210, rather than the reported 153.25. This is an arithmetic inconsistency, although the corresponding P-value remains far below .001. As an internal check on the same method, the four reported age-group counts reproduce the paper’s χ²(3) almost exactly: 47,144.9918 versus 47,145. The discrepancy is therefore specific to the sex statistic, rather than an inappropriate general choice of test.
Regression coefficients versus year-specific observed fractions. The Methods specify a modified Poisson log-link regression whose binary outcome is whether onset occurred in the birthday week, using sex, age group, birth month and reporting year. Table 2 reports an adjusted 2022-versus-2008 risk ratio of 0.792. The published marginal fractions yield an observed ratio of 0.082095, roughly one-tenth as large. Adjusted and unadjusted ratios need not agree. But if the fitted model contains only the reported categorical main effects, with no unreported interactions, additional factors, special weights or offsets, the other displayed coefficients constrain the possible marginal 2022/2008 ratio to 0.580–1.082, regardless of how the groups are composed. The observed 0.082 falls outside that mathematical range. Thus Table 2 cannot have generated the displayed marginal data under the model as described. A different, undisclosed model or outcome could reconcile the discrepancy, but the published account does not document such a difference.
Independent age-group check. The observed birthday-week fraction among infants 0–1 year is 30,539/90,709; among those older than 7 years it is 4,524/41,755, producing a crude fraction ratio of 3.107. With the displayed age effect of 1.182 and the reported sex, birth-month and year effects, the same no-interaction main-effects model constrains that crude ratio to at most 1.669. This second inequality fails independently of the calendar-year comparison.
Published code defects of limited explanatory reach. The authors’ R script includes an extra day for some birthday anniversaries crossing leap years and computes an age category one year too young for an onset exactly on the birthday. The leap-window problem is reproducible on synthetic dates. But it cannot plausibly explain the headline result: the additional day falls within a published daily bin containing only 3,308 cases, whereas the entire birthday-week numerator is 116,976. In the extreme assumption that all 3,308 were wrongly included, the fraction would remain 11.68%, still approximately six times 7/365. This is a worst-case arithmetic bound, not a patient-level correction. The publicly posted R script also does not show the regression fitting command, so Table 2 cannot be reproduced from the supplied script alone.
Scope of the regression finding. This is an incompatibility with the model as reported: a conventional, unweighted log-link model with the listed main effects only. An unreported interaction, weighting scheme, offset, different outcome, or erroneous table label could change the bound. Without the fitted model code and the individual Yunnan records, we cannot determine which explanation is correct.
2. A documented mechanism for artificial birth dates
The national Chinese infectious-disease reporting instructions issued in 2007 explicitly permitted staff to enter the patient’s age instead of the exact date of birth if the date was unknown. Comparable language appears in later official instructions.
More directly, a 2005 national China CDC reporting-system training manual, 传染病与突发公共卫生事件报告管理信息系统培训教程 (People’s Medical Publishing House; ISBN 9787117071277), describes an operational feature in section 5.1.1: for a patient with unknown birth date, enter age and the system automatically calculates the date of birth using the current date. The excerpt is available through a medical textbook transcription; an independent bibliographic catalogue confirms the existence, publisher and year of the training manual. This establishes a historically documented age-to-date conversion mechanism in that reporting-system family. It does not prove the same software or rule was used for the 2008–2022 Yunnan extract, nor does the available excerpt unambiguously specify how the day of month was assigned. Those are the decisive missing facts.
Independent evidence from a working national system (2005–2011). Wang Zhibin and Li Peng examined 225 tuberculosis-reporting cards selected from 104 groups of repeated reports concerning the same patients in Puyang, Henan. They found that 132 of the 225 cards (58.41%) had dates of birth automatically generated by the reporting system, and birth-date entries differed within all 104 repeated-report groups. Their study names the national disease surveillance reporting-management system as the source. The 58.41% figure applies only to those selected duplicate-reporting cards, not to all Chinese cases, not to children and not to the Yunnan HFMD dataset. Wang and Li, Disease Surveillance (2013).
Relevance to HFMD. Chinese national HFMD reporting used the national notifiable-disease network beginning in May 2008. Official China CDC instructions from 2009 show that records in the national disease surveillance reporting system could feed disease-specific surveillance, and that either date of birth or age could be recorded. Electronic reporting rules were subsequently revised, including in 2015. This supports continuity of the system family, not the use of one unchanging software version across 2008–2022. China CDC, 2009; Ministry of Health, 2008.
What is missing. The publicly accessible transcription of the 2005 manual does not unambiguously establish how the day of the month was calculated, nor whether that rule was applied to this particular Yunnan export. A 2013 Chinese national reporting quality survey measured agreement for a combined “birth date/age” field; that does not establish accuracy of specific calendar dates. National quality survey.
3. How a routine reporting delay could imitate a birthday effect
Falsifiable mechanism. Suppose a child whose actual birth date is unknown presents on 17 June with symptoms that began on 15 June. If a completed age of 3 years is converted to an artificial birth date of 17 June three years earlier, the recorded birthday falls exactly two days after symptom onset. The original paper’s algorithm then counts this as a disease beginning two days before the birthday, despite the child having no known birthday on 17 June. If this operation affected a sufficiently large subset of cases, ordinary diagnostic delays would manufacture a sharply rising pre-birthday curve. By contrast, a month-only or first-of-month imputation algorithm would produce a materially different pattern. The date-of-month rule matters and remains unverified.
A separate cohort from Hanzhong, Shaanxi (Wei et al., 2020, 14,814 HFMD patients) documented onset-to-hospital-visit delays: 0–1 days 56.9%, 2–3 days 34.1%, 4–7 days 7.7%, and more than seven days 1.3%. Its patients were reported online on their visit day. After subtracting the naive uniform daily background from Jiang’s seven-day counts, the excess falls into 0–1, 2–3 and 4–6 day groups at 55.9%, 36.2% and 7.9%. The external Hanzhong delay percentages restricted to delays no longer than seven days are 57.6%, 34.5% and 7.8% in corresponding 0–1, 2–3 and 4–7 day groups. This resemblance is suggestive, not a goodness-of-fit test: the final bins differ, the regions and years differ, the baseline is simplified, and the comparison was identified after inspecting the observed curve.
| Relative interval | Excess before recorded birthday in Jiang et al. | Visit-delay share in an independent HFMD cohort |
|---|---|---|
| 0–1 day | 55.9% | 57.6% |
| 2–3 days | 36.2% | 34.5% |
| 4–6 / 4–7 days | 7.9% (4–6 days) | 7.8% (4–7 days) |
Exploratory, post hoc shape comparison, not an inferential test. The last bins differ, and the baseline used to calculate excess cases is a simple 1/365 convention. These percentages cannot demonstrate that Yunnan records contained synthetic birth dates.
A simple mathematical demonstration is sufficient to show potential scale. If birthday dates were independent of onset for a fraction 1−f of recorded cases, while another fraction f had artificially reconstructed birthdays such that onset automatically appeared in the preceding week, the observed 12.02% could be reproduced by f≈10.3%. The analogous hypothetical values are approximately 32.4% for 2008 and 0.9% for 2022. These are sufficient mixture-model inputs, not measured prevalence of imputed records. This model demonstrates plausibility, not what happened in Yunnan.
4. What would confirm or falsify the explanation?
A definitive analysis requires the data dictionary and lawful, privacy-preserving review of provenance flags identifying (i) date of birth transcribed from a document or household record, (ii) age entered without a date of birth, and (iii) automatically reconstructed or defaulted birth dates. It additionally needs onset, diagnosis, visit, online reporting and age-unit fields across 2008–2022.
The strongest predictions are: age-derived dates cluster on the calendar day of diagnosis/reporting (or on a predictable alternative day-of-month rule); the apparent days-before-birthday distribution tracks onset-to-reporting lag among age-derived records; and the excess substantially declines when independently verified birth dates are analyzed alone. Its calendar-year trend should covary with the share of age-only entries. These tests should first establish the exact transformation rule, and only then estimate how much of the sixfold association survives.
If the Yunnan record dates were all independently verified against birth documents and remain uncorrelated with reporting dates, this specific date-generation hypothesis would be falsified. The chi-square and regression-consistency problems would remain, but would not by themselves disprove a real biological pattern.
Conclusion: three different levels of certainty
Reproduced: a discrepancy in the reported sex chi-square statistic; bounds showing that the year and age margins cannot be generated by the listed no-interaction Poisson main-effects model as described; synthetic-date bugs in the publicly released R script. These are checks of published calculations and specified methods, not a reproduction of the original authors’ patient-level analysis.
Documented but not linked to the cohort: the historical Chinese national disease-reporting system could automatically generate birth dates from ages, and this happened in a subset of real, selected reporting cards during 2005–2011.
Unresolved: whether Yunnan’s 2008–2022 HFMD data contained these inferred dates; how the numeric day-of-month was set; what fraction of the sixfold cluster would remain among document-verified birth dates. The sixfold pattern is real in the published table, but its interpretation as an increase in disease incidence or susceptibility is not established by the public evidence reviewed here.
Research status. This is a public institutional methodological note, not a peer-reviewed correction, a retraction request, or a verified alternative estimate of HFMD incidence. We have not asked the authors for confidential data. We would welcome disclosure of the original DOB-provenance data dictionary and the exact regression-fitting script, along with non-identifying analyses separating documented and age-derived birthdays.
Sources and computational reproducibility
- Jiang et al. (2024), JMIR Public Health and Surveillance: original study and Tables 1–2, Figure 1.
- Authors’ public R script. Examined version SHA-256:
f8ea7cf98f1bec84074050b249c19579a6fcd3036214aaf5a6134cb0ad88e9cf. - Wang and Li (2013), Disease Surveillance: duplicate tuberculosis reporting and automatically generated birth dates, 2005–2011.
- 2005 national disease-reporting software manual, transcribed section 5.1.1; independent bibliographic entry. The original printed page was not inspected.
- China CDC 2009 reporting instructions; HFMD reporting guidance, 2008; 2015 reporting standards.
- Wei et al. (2020): independent Chinese HFMD reporting-delay distribution.
- China CDC 2013 reporting-card quality survey.
Reproducible arithmetic: Pearson’s χ² can be calculated from the 2 × 2 table [68,849, 495,876; 48,127, 360,558]. The regression compatibility check bounds the 2022-to-2008 marginal ratio by multiplying its reported year RR (0.792) by the minimum and maximum possible ratios of the other published main-effect factors. Both checks require only the numbers in Jiang et al.’s tables. Individual Yunnan records, the original modified-Poisson fitting code, and provenance flags were not publicly available to this audit.