Evidence level: B
三方头对头真正难的不是”谁赢”,而是三方的缺失机制各不相同——若把三方硬塞进一个模型,”有效样本”会自动落在缺失最不坏的那一小块上,胜负可能由缺失方向而非信息量决定。
正文
这场比赛的前两方,成绩单已经完整
容易被忽略的是:这个三方头对头里只有一方是真正未知的。家族史与多基因风险评分(PRS)的对比已在大样本里做过三轮。
第一轮是 FinnGen(N=306,418)对 24 种常见病的系统对比(PMID 36347255):家族史与 PRS 独立且不可互换——PRS 平均只解释家族史效应的 10%,家族史只解释 PRS 效应的 3%。第二轮是 UK Biobank(约 40 万欧洲血统参与者)的联合预测器(PMID 35710731):心肌梗死上联合模型 AUROC 0.7625,单用 PRS 0.7567,单用父母病史 0.7375。第三轮是冠心病的加性检验(PMID 38939477):323,373 名参与者中,PRS 加入池化队列方程使 AUC 从 0.759 升到 0.773,且 PRS、家族性高胆固醇血症、家族史三者风险独立且可乘性相加。
三轮合起来说明:这两方的判别力都落在 0.75-0.78,各自贡献的增量在 0.01-0.02 量级。这不是弱,是复杂疾病的常态。
第三方的位置,也已被间接测出
出生时刻没被放进过同一队列,但它的量级已被单独测过多次,而且测出来的数字有一半是假象。
先看最大的那些:瑞典 210 万例中,夜间(21:00-09:00)出生的早产儿早期新生儿死亡相对风险 1.31、窒息相关死亡 1.70(PMID 11739216);瑞典 694,888 例单胎中,夜间出生早期新生儿死亡调整后 RR 1.28(PMID 12606889);奥地利 87,065 例中,夜间严重不良事件 OR 1.35(PMID 29210161)。这些数字比季节级效应(OR≈1.05)大一个量级——单看这里,会以为”时辰”确实携带了远超”季节”的信息。
但冷水马上浇下来。英格兰 6,054,536 例的链接队列研究在校正混杂后,工作时段与非工作时段的死亡几率没有显著差异;研究者进一步发现,那个”看似存在的周末效应”其实来自极少数”无产程记录的急诊剖宫产”(PMID 37311636)。美国 NICHD 的 11,137 例极低出生体重婴儿队列更直接:出生时机对死亡率与短期并发症没有影响(PMID 20643715)。韩国 3,220 例也只在 NICU 水平较低的医院组里看到夜间效应(PMID 33821593)。
换句话说,时辰级”效应”在人力充足时可以被管理措施消掉。能被排班消掉的东西,归属的是排班,不是天时。
更根本的问题是:出生时刻的分布本身被机构日程重塑。Declercq 等(2023)对比美、英、荷三国,家庭分娩峰值稳定落在凌晨 2:00-5:00,英荷的医院自发产保持类似形态,而美国医院的自发产峰值反转为 8:00-17:00(PMID 36652413)。作者原话是”组织性优先事项可能扰动自然的产程与出生时机模式”。在美国这类数据里做时辰分析,测到的可能是日程表。
真正难在哪:三方缺失机制不同
到这一步,”谁赢”其实已可预判:时柱项的判别增量大概率落在小数点后第三位,而家族史与 PRS 落在第二位。真正的学术价值不在这个排序,而在这个对比能否被公平地做出来。
难点是三方各自的”缺失”不是同一种缺失:家族史的缺失是”不知道“(与年龄、家庭联系紧密度相关);PRS 的缺失是”没测“(多由生物银行设计决定);出生时刻的缺失是”没记“,而这一项恰恰最不随机——英格兰那项研究正是通过缺失模式发现,出生时刻记录不全的那批人集中在”无产程记录的急诊剖宫产”里(PMID 37311636)。
把三方塞进同一个模型,实际进入分析的样本必然是三方都齐全的交集。由于出生时刻的缺失与分娩方式相关、分娩方式又与结局相关,这个交集会把时辰项的判别力系统性地偏向某一侧。若不做这一步处理,报出来的”时辰无效”可能是缺失造成的,”时辰有效”也可能是缺失造成的。
指标不能靠单一数字裁决
第二个坑是评价指标本身。Pepe 等证明:即使新增标记毫无预测信息,在训练集上算出的净重分类指数(NRI)也倾向为正(PMID 26504496);四个无预测能力的标记重复模拟下,NRI 假阳性率训练集高达 63.0%、测试集 18.8%-34.4%,而 ΔAUC 的假阳性率维持在预期的 5% 附近(PMID 24681599)。Kerr 等的综述建议:三档以上风险分类时不要用 NRI,首选单数字摘要应是净获益的改善(PMID 24240655);即使用,置信区间也应走 bootstrap(PMID 26891393)。
所以判据不能是”谁的 NRI 大”,合理做法是三指标同报:ΔAUC + 净获益 + 校准,并事先写下增量可判的门槛。另需澄清一个误解:ΔAUC 小不等于无价值——Martens 等发现 ΔAUC 0.01 时预测风险变化确实也小,但 AUC 并非像常说的那样不敏感(PMID 27430730);稀有变异的所有模拟场景中 ΔAUC 上限都没超过 0.05(PMID 23961719)。“小时辰”要论证自己有价值,必须走净获益,不能只靠 AUC 或 NRI 的话术。
易学落点:四柱里的”时”,本来就不该与”祖荫”同权
八字把年月日时四柱当作四个并列坐标。但传统内部一直混着两种来源不同的东西:一类是当事人出生那一瞬的天时(时柱、日躔),另一类是家族传递下来的东西(祖上风水、父母根气、祖荫)。古人从未把这两类当作同一回事——这恰与现代遗传学结论同构:家族史与 PRS 独立、互补、不可互换(PMID 36347255)。
分歧在权重。传统默认”时柱”与”祖荫”同量级;而现有数据显示,家族史与 PRS 的个体判别增量在 0.01-0.02 量级,时辰级的直接效应则在控制排班与人力后趋近于零(PMID 37311636、PMID 20643715)。
可证伪的预测:若在同一队列里做三方头对头,控制机构排班变量(引产率、夜间人力、分娩方式)后,时柱项的 ΔAUC 应显著小于家族史项与 PRS 项,且其”看似有效”的部分应大部分被排班变量解释。若时柱项在控制这些变量后仍有稳定的净获益增量,那才是这个体系真正需要解释的现象。## 证据与局限
已验证: 家族史与 PRS 在 24 种常见病中独立且不可互换,PRS 仅解释家族史效应的 10%、家族史仅解释 PRS 效应的 3%(PMID 36347255);UK Biobank 联合预测器在心肌梗死上 AUROC 0.7625 vs 单 PRS 0.7567 vs 单父母病史 0.7375(PMID 35710731);PRS、家族性高胆固醇血症、家族史对冠心病的风险独立且可乘性相加(PMID 38939477);时辰级出生效应的最大报告值在 RR 1.28-1.35 量级(PMID 11739216、12606889、29210161),但校正混杂后归零(PMID 37311636)或在人力充足队列中无影响(PMID 20643715、33821593);出生时刻分布被机构日程重塑,美国医院自发产峰值反转为 8:00-17:00 而家庭分娩稳定在 2:00-5:00(PMID 36652413);NRI 在无信息标记上倾向为正、假阳性率训练集 63.0%(PMID 26504496、24681599);ΔAUC 与 IDI 随基线 AUC 升高而降低(PMID 27430730),稀有变异 ΔAUC 上限 <0.05(PMID 23961719)。
部分验证: 出生时刻的记录缺失与分娩方式相关(英格兰队列从缺失模式反推出”周末效应”来自无产程记录的急诊剖宫产,PMID 37311636);家族史与 PRS 的判别增量均在 0.01-0.02 量级,但两者在不同疾病上的具体排序随疾病遗传结构而变,尚无统一结论。
待验证: ① 同一队列内”出生时刻(时辰粒度)×家族史×PRS”三方判别力的头对头正式研究——尚无文献在同一时间同一批样本中同时测量三者的 ΔAUC/净获益;② 三方缺失机制不同所造成的”交集样本选择效应”从未被量化,无法判断时辰项判别力的偏移方向与幅度;③ 时辰项在控制机构排班变量(引产率、夜间人力、分娩方式)后的残余判别力从未被测量。
相关阅读
- 出生季节效应 vs 家族史/多基因风险评分:个体判别力的头对头对比
- 出生时刻(时辰)粒度的可判别信息量上界
- 出生月份信息量鸿沟的量化
- 命理”判定准确率门槛”的可操作定义与预注册
- 研究池子:时空频率拓扑
证据等级: B(科学类比/间接证据)
参考文献: PMID 36347255, PMID 35710731, PMID 38939477, PMID 11739216, PMID 12606889, PMID 29210161, PMID 37311636, PMID 20643715, PMID 33821593, PMID 36652413, PMID 26504496, PMID 24681599, PMID 24240655, PMID 26891393, PMID 27430730, PMID 23961719
免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。
Birth Hour × Family History × Polygenic Risk Score: A Head-to-Head Design of Tripartite Individual Discriminatory Power
Evidence level: B
The real difficulty in this three-way head-to-head is not “who wins,” but that each side has distinct missing mechanisms. If forced into a single model, the “valid sample” will automatically fall onto the smallest subset with the least problematic missingness; thus, the outcome may be determined by the direction of missingness rather than information content.
Main Text
The First Two Sides Already Have Complete Scorecards
What is often overlooked is that only one side in this three-way head-to-head is truly unknown. Comparisons between family history and Polygenic Risk Score (PRS) have already been conducted in large samples across three rounds.
Round one was a systematic comparison in FinnGen (N=306,418) against 24 common conditions (PMID 36347255): Family history and PRS are independent and non-interchangeable—PRS explains only about 10% of the family history effect, while family history explains only 3% of the PRS effect. Round two was a joint predictor in UK Biobank (approx. 400,000 participants of European ancestry) (PMID 35710731): For myocardial infarction, the joint model achieved an AUROC of 0.7625, compared to 0.7567 for PRS alone and 0.7375 for parental history alone. Round three was an additive test for coronary heart disease (PMID 38939477): Among 323,373 participants, adding PRS to the pooled cohort equation increased AUC from 0.759 to 0.773, and PRS, familial hypercholesterolemia, and family history showed independent and multiplicatively additive risks.
Taken together, these three rounds indicate that the discriminative power of both sides falls within the 0.75–0.78 range, with incremental contributions at the 0.01–0.02 magnitude. This is not weakness; it is the norm for complex diseases.
The Position of the Third Side Has Also Been Indirectly Measured
Birth time has not been included in the same cohort, but its magnitude has been measured multiple times separately, and half of the resulting numbers are artifacts.
First, look at the largest datasets: Among 2.1 million cases in Sweden, preterm infants born at night (21:00–09:00) had a relative risk of early neonatal death of 1.31 and suffocation-related deaths of 1.70 (PMID 11739216); among 694,888 singletons in Sweden, the adjusted RR for early neonatal death was 1.28 for night births (PMID 12606889); and among 87,065 cases in Austria, the OR for severe adverse events was 1.35 for night births (PMID 29210161). These figures are an order of magnitude larger than seasonal effects (OR≈1.05)—looking at this alone, one might assume “birth time” carries information far exceeding that of “season.”
But cold water is immediately poured on this. In a linked cohort study of 6,054,536 cases in England, after adjusting for confounders, there was no significant difference in mortality odds between working hours and non-working hours. Researchers further found that the “apparent weekend effect” actually stemmed from a small number of “emergency cesarean sections without labor records” (PMID 37311636). The US NICHD cohort of 11,137 extremely low birth weight infants was even more direct: timing of birth had no effect on mortality and short-term complications (PMID 20643715). In Korea, the night effect was observed only in the low-level NICU hospital group among 3,220 cases (PMID 33821593).
In short, time-of-day-level “effects” can be eliminated by management measures when staffing is sufficient. Things that can be eliminated by scheduling belong to the schedule, not the celestial timing.
A more fundamental issue is that the distribution of birth moments itself is reshaped by institutional schedules. Declercq et al. (2023) compared the US, UK, and Netherlands; family birth peaks remained stable between 2:00–5:00 AM, and spontaneous births in UK/Netherlands hospitals maintained a similar shape, whereas the peak for spontaneous births in US hospitals reversed to 8:00–17:00 (PMID 36652413). The authors stated, “Organizational priorities may disrupt natural labor and birth timing patterns.” In such US data, time-of-day analysis might merely be measuring the schedule.
The Real Difficulty: Different Missing Mechanisms Across Three Sides
At this stage, the outcome of “who wins” can be anticipated: the discriminative increment for the time pillar likely falls in the third decimal place, while family history and PRS fall in the second. The true academic value lies not in this ranking, but in whether this comparison can be conducted fairly.
The difficulty is that the “missingness” for each side is not the same: Family history missingness is “unknown” (correlated with age and family contact density); PRS missingness is “not tested” (mostly determined by biobank design); and birth time missingness is “not recorded,” which is precisely the least random—the English study identified via missing patterns that those with incomplete birth time records concentrated in “emergency cesarean sections without labor records” (PMID 37311636).
Forcing all three sides into a single model means the actual sample entering analysis must be the intersection where all three are complete. Since birth time missingness correlates with delivery mode, and delivery mode correlates with outcomes, this intersection will systematically bias the discriminative power of the time pillar. Without addressing this step, reported “ineffectiveness” of time may be caused by missingness, as might “effectiveness.”
Metrics Cannot Be Decided by a Single Number
The second pitfall is the metric itself. Pepe et al. proved that even if new markers have no predictive information, the Net Reclassification Index (NRI) calculated on the training set tends to be positive (PMID 26504496); under four simulations of markers with no predictive ability, the NRI false positive rate was 63.0% on the training set and 18.8%–34.4% on the test set, while the false positive rate for ΔAUC remained near the expected 5% (PMID 24681599). Kerr et al.’s review suggests not using NRI for risk stratification with three or more categories; the preferred single-number summary should be improvement in net benefit (PMID 24240655); if used, confidence intervals should be bootstrapped (PMID 26891393).
Therefore, the criterion cannot be “whose NRI is larger.” A reasonable approach is to report all three metrics simultaneously: ΔAUC + Net Benefit + Calibration, and pre-specify thresholds for incremental significance. Another misconception must be clarified: small ΔAUC does not equal no value—Martens et al. found that while prediction risk changes were indeed small at ΔAUC 0.01, AUC is not insensitive as often claimed (PMID 27430730); the upper limit for ΔAUC in all simulation scenarios for rare variants did not exceed 0.05 (PMID 23961719). “Small time” must demonstrate value through net benefit; relying solely on AUC or NRI rhetoric is insufficient.
Yijing Focus: The “Hour” in Four Pillars Should Not Share Equal Weight with “Ancestral Favor”
Bazi treats the four pillars of year, month, day, and hour as four parallel coordinates. However, traditional internal logic mixes two distinct sources: one is the celestial timing at the instant of birth (hour pillar, day position), and the other is inherited family elements (ancestral Feng Shui, parental root energy, ancestral favor). The ancients never treated these two categories as identical—this aligns structurally with modern genetic conclusions: family history and PRS are independent, complementary, and non-interchangeable (PMID 36347255).
The divergence lies in weighting. Tradition defaults to equating the “hour pillar” with “ancestral favor” in magnitude; however, existing data shows that the individual discriminative increments for family history and PRS fall in the 0.01–0.02 magnitude, whereas the direct effect of time-of-day approaches zero after controlling for staffing and labor patterns (PMID 37311636, PMID 20643715).
Falsifiable Prediction: If a three-way head-to-head is conducted within the same cohort, controlling for institutional scheduling variables (induction rates, night staffing, delivery mode), the ΔAUC for the time pillar should be significantly smaller than that for family history and PRS, and its “apparently effective” portion should be largely explained by scheduling variables. Only if the time pillar shows a stable net benefit increment after controlling for these variables represents a phenomenon this system truly needs to explain.
Evidence and Limitations
Verified: Family history and PRS are independent and non-interchangeable across 24 common conditions; PRS explains only 10% of the family history effect, and family history explains only 3% of the PRS effect (PMID 36347255); UK Biobank joint predictor achieved AUROC 0.7625 for myocardial infarction vs. 0.7567 for single PRS and 0.7375 for single parental history (PMID 35710731); PRS, familial hypercholesterolemia, and family history show independent and multiplicatively additive risks for coronary heart disease (PMID 38939477); maximum reported values for time-of-day birth effects fall in the RR 1.28–1.35 magnitude (PMID 11739216, 12606889, 29210161), but vanish after adjusting for confounders (PMID 37311636) or show no effect in cohorts with sufficient staffing (PMID 20643715, 33821593); birth time distribution is reshaped by institutional schedules, with US hospital spontaneous birth peaks reversing to 8:00–17:00 while family births remain stable at 2:00–5:00 (PMID 36652413); NRI tends to be positive for markers with no information and has a false positive rate of 63.0% on the training set (PMID 26504496, 24681599); ΔAUC and IDI decrease as baseline AUC increases (PMID 27430730), with rare variant ΔAUC upper limits <0.05 (PMID 23961719).
Partially Verified: Birth time record missingness correlates with delivery mode (English cohort inferred “weekend effect” from missing patterns to stem from emergency cesareans without labor records, PMID 37311636); discriminative increments for family history and PRS both fall in the 0.01–0.02 magnitude, but their specific rankings across different diseases vary with genetic structure, with no unified conclusion.
To Be Verified: ① Formal head-to-head studies of three-way discriminative power within the same cohort for “birth time (hour granularity) × family history × PRS”—no literature simultaneously measures the ΔAUC/net benefit of all three in the same time and sample; ② The “intersection sample selection effect” caused by different missing mechanisms across three sides has never been quantified, making it impossible to judge the direction and magnitude of bias in time pillar discriminative power; ③ Residual discriminative power of the time pillar after controlling for institutional scheduling variables (induction rates, night staffing, delivery mode) has never been measured.
Related Reading
- Birth Season Effect vs Family History/Polygenic Risk Score: Head-to-Head Comparison of Individual Discriminative Power
- Upper Bound of Discriminative Information at Birth Time (Hour) Granularity
- Quantification of Information Gap at Birth Month
- Operational Definition and Preregistration for Divination “Judgment Accuracy Threshold”
- Research Pool: Spatiotemporal Frequency Topology
Evidence Level: B (Scientific Analogy / Indirect Evidence)
References: PMID 36347255, PMID 35710731, PMID 38939477, PMID 11739216, PMID 12606889, PMID 29210161, PMID 37311636, PMID 20643715, PMID 33821593, PMID 36652413, PMID 26504496, PMID 24681599, PMID 24240655, PMID 26891393, PMID 27430730, PMID 23961719
Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.
Leave a Reply