打這支疫苗竟可降癡呆風險？研究揭示「非預期」保護效應！

本翻譯僅作學術交流用，無商業意圖，請勿轉載，如有疑議問請來信

英國研究發現，接種活減毒帶狀皰疹疫苗（Zostavax）者，7 年內罹患癡呆的機率下降約 20%。透過自然實驗設計，研究排除多數混淆因素，首次提供此疫苗可能延緩癡呆發生的因果證據，女性保護效益更顯著。

一項關於帶狀皰疹疫苗對癡呆症影響的自然實驗

A natural experiment on the effect of herpes zoster vaccination on dementia

Eyting M, Xie M, Michalik F, Heß S, Chung S, Geldsetzer P. A natural experiment on the effect of herpes zoster vaccination on dementia. Nature. 2025;641(8062):438-446. doi:10.1038/s41586-025-08800-x

https://pubmed.ncbi.nlm.nih.gov/40175543/

摘要 Abstract

神經性疱疹病毒可能與癡呆的發展有關 ^1,2,3,4,5 。此外，疫苗可能具有重要的非靶向免疫學效應 ^6,7,8,9 。我們的目標是確定活減毒帶狀皰疹疫苗對癡呆診斷發生的影響。為了提供因果而非相關的證據，我們利用威爾士的事實，即帶狀皰疹疫苗的資格是根據個人的確切出生日期來決定的。1933 年 9 月 2 日之前出生的人終身不符合資格，而 1933 年 9 月 2 日或之後出生的人至少有 1 年的時間可以接種疫苗。利用大規模電子健康記錄數據，我們首先顯示，接種疫苗的成年人比例從不符合資格的患者中僅為 0.01%，增加到剛好年輕 1 週的患者中達到 47.2%。除了這種接種帶狀皰疹疫苗的概率的巨大差異外，1933 年 9 月 2 日之前出生的個體與 1933 年 9 月 2 日之後出生的個體在系統上不太可能存在差異。 使用這些比較組別進行回歸不連續性設計，我們顯示接種帶狀皰疹疫苗在 7 年的隨訪期間降低了新診斷癡呆症的概率 3.5 個百分點（95%置信區間（CI）= 0.6–7.1，P = 0.019），相當於 20.0%（95% CI = 6.5–33.4）的相對減少。這種保護效果在女性中比男性更強。我們在不同的人群（英格蘭和威爾士的綜合人群）中成功確認了我們的發現，使用不同類型的數據（死亡證明）和一個與癡呆症密切相關但不太依賴於醫療系統及時診斷癡呆症的結果（以癡呆症為主要原因的死亡） ¹⁰ 。通過使用獨特的自然實驗，本研究提供了帶狀皰疹疫苗對癡呆症的預防或延遲效果的證據，這比現有的關聯證據更不易受到混淆和偏見的影響。

Neurotropic herpesviruses may be implicated in the development of dementia^1,2,3,4,5. Moreover, vaccines may have important off-target immunological effects^6,7,8,9. Here we aim to determine the effect of live-attenuated herpes zoster vaccination on the occurrence of dementia diagnoses. To provide causal as opposed to correlational evidence, we take advantage of the fact that, in Wales, eligibility for the zoster vaccine was determined on the basis of an individual’s exact date of birth. Those born before 2 September 1933 were ineligible and remained ineligible for life, whereas those born on or after 2 September 1933 were eligible for at least 1 year to receive the vaccine. Using large-scale electronic health record data, we first show that the percentage of adults who received the vaccine increased from 0.01% among patients who were merely 1 week too old to be eligible, to 47.2% among those who were just 1 week younger. Apart from this large difference in the probability of ever receiving the zoster vaccine, individuals born just 1 week before 2 September 1933 are unlikely to differ systematically from those born 1 week later. Using these comparison groups in a regression discontinuity design, we show that receiving the zoster vaccine reduced the probability of a new dementia diagnosis over a follow-up period of 7 years by 3.5 percentage points (95% confidence interval (CI) = 0.6–7.1, P = 0.019), corresponding to a 20.0% (95% CI = 6.5–33.4) relative reduction. This protective effect was stronger among women than men. We successfully confirm our findings in a different population (England and Wales’s combined population), with a different type of data (death certificates) and using an outcome (deaths with dementia as primary cause) that is closely related to dementia, but less reliant on a timely diagnosis of dementia by the healthcare system¹⁰. Through the use of a unique natural experiment, this study provides evidence of a dementia-preventing or dementia-delaying effect from zoster vaccination that is less vulnerable to confounding and bias than the existing associational evidence.

主要 Main

最近，有證據顯示神經嗜好性單純皰疹病毒可能在癡呆的發病機制中扮演角色 ^1,2,3,4,5 。針對單純皰疹病毒的一種方法是接種疫苗。然而，疫苗也越來越被認識到能引發更廣泛的免疫反應，這可能會產生重要的非靶向效應，特別是在活減毒疫苗的情況下 ^6,7,8,9 。這種效應經常被觀察到在性別上有明顯的差異 ⁷ 。

Recently, evidence has grown that neurotropic herpesviruses may have a role in the pathogenesis of dementia^1,2,3,4,5. One approach to targeting herpesviruses is vaccination. However, vaccines are also increasingly being recognized as eliciting a broader immune response that can have important off-target effects, particularly in the case of live-attenuated vaccines^6,7,8,9. Such effects have frequently been observed to differ strongly by sex⁷.

截至目前，對於疫苗接種對癡呆症影響的隊列研究和電子健康記錄數據僅僅比較了接種特定疫苗的人與未接種者之間癡呆症的發生率 ¹¹ 。這些研究必須假設接種者與未接種者之間所有不同的特徵（且這些特徵也與癡呆症相關）在分析中已經得到了充分的測量和建模，以至於沒有因素會混淆疫苗接種與癡呆症之間的關係 ¹² 。這種沒有混淆偏差的假設往往是不可信的，因為必須假設研究擁有關於難以測量的因素（如個人動機或健康素養）的詳細數據 ¹³ 。這也是一個無法經驗驗證的假設。

To date, studies in cohort and electronic health record data on the effect of vaccination receipt on dementia have simply compared the occurrence of dementia among those who received a given vaccination and those who did not¹¹. These studies have to assume that all characteristics that are different between those who are vaccinated and those who are not (and that are also related to dementia) have been sufficiently well measured and modelled in the analysis, such that no factors confound the relationship between vaccination receipt and dementia¹². This assumption of no confounding bias is often implausible because it has to be assumed that the study has detailed data on factors that are difficult to measure, such as personal motivation or health literacy¹³. It is also an assumption that cannot be empirically verified.

我們採用了根本不同的方法，利用了這個事實：在威爾士，自 2013 年 9 月 1 日起，1933 年 9 月 2 日或之後出生的人有資格接種帶狀皰疹疫苗至少一年，而早於此日期出生的人則永遠無法獲得資格 ¹⁴ 。通過使用詳細的大規模電子健康記錄數據，我們能夠比較那些因為出生在資格截止日期之前而無法接種疫苗的成年人與那些在資格截止日期之後出生且有資格接種的人。重要的是，年齡相差僅幾週的個體預期不會系統性地彼此不同。也就是說，我們的比較組之間所有潛在的混淆變量在預期上是平衡的。通過利用這一獨特的自然實驗，我們能夠比所有現有的相關研究更可信地避免混淆 ^{15,16,17,18,19,20,21,22,23,24} ，這些研究僅僅是將疫苗接種者與未接種者進行比較，同時試圖控制這些組之間的各種差異。

We used a fundamentally different approach that takes advantage of the fact that, in Wales, starting on 1 September 2013, those born on or after 2 September 1933 were eligible for herpes zoster vaccination for at least 1 year, while those born earlier never became eligible¹⁴. Using detailed large-scale electronic health record data, we were able to compare adults who were ineligible for the vaccine because they were born immediately before the eligibility cut-off date with those born immediately after who were eligible. Importantly, individuals who are only a few weeks apart in age are not expected to differ systematically from each other. That is, all potential confounding variables are in expectation balanced between our comparison groups. By taking advantage of this unique natural experiment, we were able to avoid confounding more credibly than all existing studies on the topic^{15,16,17,18,19,20,21,22,23,24}, which have simply compared vaccine recipients to non-recipients while trying to control for the myriad of differences between these groups.

在 1933 年 9 月 2 日的出生資格截止日期之後出生的成年人，接種帶狀皰疹疫苗的概率比在此截止日期之前出生的人高出 47.2 個百分點（從 0.01%到 47.2%）。如預期的那樣，除了帶狀皰疹疫苗接種率的這一突變外，患者在 1933 年 9 月 2 日的出生資格門檻上，在其他預防健康服務的使用、過去的常見疾病診斷和教育程度方面是平衡的。然後，我們在回歸不連續性分析中使用這一“準隨機化”方法，首先重複了臨床試驗中已知的發現，即帶狀皰疹疫苗減少了新診斷的帶狀皰疹。其次，我們將這一方法擴展到一個在帶狀皰疹疫苗的臨床試驗中從未評估的結果——癡呆，發現該疫苗在七年的隨訪期間將新癡呆診斷的概率降低了約五分之一。第三，我們顯示帶狀皰疹疫苗並未影響除帶狀皰疹和癡呆之外的任何其他常見死亡或發病原因的發生。 同樣地，我們顯示接種帶狀皰疹疫苗並未導致其他疫苗或預防健康措施的增加。第四，我們提供證據顯示，在威爾士，沒有其他干預措施（例如健康保險資格）使用與帶狀皰疹疫苗相同的出生日期（1933 年 9 月 2 日）作為資格截止日期。第五，我們顯示當使用不同的分析方法時，所有發現仍然相似。第六，我們顯示由於帶狀皰疹發作而導致的醫療途徑變化不太可能解釋我們的發現。第七，我們提供來自電子健康記錄數據的探索性證據，說明帶狀皰疹疫苗如何影響癡呆症的機制。我們的研究專注於活減毒帶狀皰疹疫苗（Zostavax；以下簡稱帶狀皰疹疫苗），因為較新的重組亞單位帶狀皰疹疫苗（Shingrix）在我們的隨訪期間結束後才在英國上市。

Adults born immediately after the 2 September 1933 date-of-birth eligibility cut-off had a 47.2 percentage point higher probability (from 0.01% to 47.2%) of ever receiving the herpes zoster vaccine than those born immediately before this cut-off date. As expected, other than this abrupt change in herpes zoster vaccination uptake, patients were balanced across the 2 September 1933 date-of-birth eligibility threshold in their uptake of other preventive health services, past common disease diagnoses and educational attainment. We then used this ‘quasi-randomization’ in a regression discontinuity analysis to first replicate the known finding from clinical trials that the herpes zoster vaccine reduces new diagnoses of shingles. Second, we extended this approach to an outcome—dementia—that was never assessed in clinical trials of the herpes zoster vaccine, and find that the vaccine reduces the probability of a new dementia diagnosis over a seven-year follow-up period by approximately one-fifth. Third, we show that the herpes zoster vaccine did not affect the occurrence of any other common causes of mortality or morbidity other than shingles and dementia. Similarly, we show that receipt of the herpes zoster vaccine did not lead to increased uptake of other vaccinations or preventive health measures. Fourth, we provide evidence that no other intervention (such as health insurance eligibility) in Wales used the identical date of birth (2 September 1933) as eligibility cut-off as was used to define eligibility for the herpes zoster vaccine. Fifth, we show that all findings remain similar when using a different analysis approach. Sixth, we show that changes in healthcare pathways as a result of a shingles episode are unlikely to explain our findings. Seventh, we provide exploratory evidence from our electronic health record data on the mechanism through which herpes zoster vaccination could affect dementia. Our study focuses on the live-attenuated herpes zoster vaccine (Zostavax; hereafter, zoster vaccine), because the newer recombinant subunit zoster vaccine (Shingrix) became available in the UK only after our follow-up period ended²⁵.

帶狀皰疹疫苗接種的差異
Difference in zoster vaccination receipt

我們使用了安全匿名信息鏈接（SAIL）數據庫 ²⁶ ，該數據庫包含來自威爾士約 80%的初級護理提供者的詳細電子健康記錄數據，並與次級護理記錄和該國的死亡登記數據相連接。我們主要分析的研究人群由所有在 1925 年 9 月 1 日至 1942 年 9 月 1 日之間出生的成年人組成，他們在威爾士的初級護理提供者處註冊（這對於居住在威爾士的成年人來說超過 98% ²⁷ ），並且在威爾士帶狀皰疹疫苗計劃開始時（2013 年 9 月 1 日）沒有被診斷為癡呆。我們主要分析隊列中 282,541 名成年人的基本社會人口學和臨床特徵顯示在補充表 1 中。

We used the Secure Anonymised Information Linkage (SAIL) Databank²⁶, which contains detailed electronic health record data on primary care visits from approximately 80% of primary care providers in Wales, linked to secondary care records and the country’s death register data. The study population for our primary analyses consisted of all adults born between 1 September 1925 and 1 September 1942 who were registered with a primary care provider (which is the case for over 98% of adults residing in Wales²⁷), resided in Wales and did not have a diagnosis of dementia at the time of the start of the zoster vaccine program in Wales (on 1 September 2013). Basic sociodemographic and clinical characteristics of the sample of 282,541 adults in our primary analysis cohort are shown in Supplementary Table 1.

在威爾斯，1933 年 9 月 2 日至 1934 年 9 月 1 日出生的個體（我們數據中的 16,595 名成年人）於 2013 年 9 月 1 日開始有資格接種帶狀皰疹疫苗，至少為期 1 年。隨後，根據他們的出生日期，年齡較小的群體每年逐步獲得資格，但年齡較大的群體則不會（方法）。

In Wales, individuals born between 2 September 1933 and 1 September 1934 (16,595 adults in our data) became eligible for the zoster vaccine for at least 1 year on 1 September 2013. Eligibility was then progressively extended to younger, but not older, age cohorts annually on the basis of their exact date of birth (Methods).

我們發現，出生於 1933 年 9 月 2 日之後僅 1 週的人，因此至少有資格接種帶狀皰疹疫苗，導致接種帶狀皰疹疫苗的概率從 0.01%急劇上升至 47.2%（P < 0.001；圖 1）。這提供了一個獨特的機會來避免混淆問題，因為在出生日期資格門檻附近出生的個體不太可能在除了年齡差異一週和接種帶狀皰疹疫苗的概率大差異之外，系統性地彼此不同。我們通過實證證明這一點，顯示在帶狀皰疹疫苗接種計劃開始時，無論是常見疾病診斷的流行率（包括在疫苗接種計劃推出之前被診斷為癡呆症），還是根據一系列臨床和社會人口變量預測的癡呆風險，亦或是預防行為的流行率（除了接種帶狀皰疹疫苗）在帶狀皰疹疫苗的出生日期資格門檻上均未顯示出不連續性（圖 1 和補充圖 1–4）。 因此，在靈活控制年齡後，我們的兩個比較組（一個是低概率接種帶狀皰疹疫苗，另一個是高概率接種帶狀皰疹疫苗）在 1933 年 9 月 2 日出生資格門檻的兩側出生，可能在所有觀察到的和未觀察到的潛在混淆變數上是可互換的。

We find that being born just 1 week after 2 September 1933, and therefore being eligible for the zoster vaccine for at least 1 year, caused an abrupt increase in the probability of ever receiving the zoster vaccine from 0.01% to 47.2% (P < 0.001; Fig. 1). This provides a unique opportunity to avoid confounding concerns because it is unlikely that individuals born immediately around the date-of-birth eligibility threshold systematically differ from each other by anything but a one-week difference in age and a large difference in the probability of receiving the zoster vaccine. We substantiate this empirically by showing that, at the time of the start date of the zoster vaccination program, neither the prevalence of common disease diagnoses (including having been diagnosed with dementia before the vaccination program rollout), dementia risk as predicted from a series of clinical and sociodemographic variables, nor the prevalence of preventive behaviours (other than zoster vaccine uptake) display a discontinuity at the date-of-birth eligibility threshold for the zoster vaccine (Fig. 1 and Supplementary Figs. 1–4). Thus, after flexibly controlling for age, our two comparison groups (one with a low and one with a high probability of receiving the zoster vaccine) born immediately on either side of the 2 September 1933 date-of-birth eligibility threshold are probably exchangeable with each other on all observed and unobserved potential confounding variables.

圖 1：在出生日期符合資格的門檻上，帶來帶狀皰疹疫苗接種的大幅增加。

a–f，出生日期的資格截止點導致帶狀皰疹疫苗接種的巨大不連續性（a），但在其他預防性干預措施的採用上（流感疫苗（d）、肺炎球菌多醣疫苗（PPV）（e）和他汀類藥物（f））在截止點上存在基線可交換性，以及過去的帶狀皰疹（b）和癡呆（c）診斷。這項分析的數據來源是威爾士的 SAIL 數據庫。所有分析均在與帶狀皰疹疫苗對癡呆發生影響的樣本相同的樣本上進行。例外的是 c，對於該項，我們並未排除在帶狀皰疹疫苗計劃開始之前已被診斷為癡呆的個體。灰色點顯示每 10 週增量的平均值。灰色點的陰影與該 10 週增量在分析中所獲得的觀察權重成比例。

a–f, The date-of-birth eligibility cut-off led to a large discontinuity in zoster vaccine receipt (a), but there is baseline exchangeability across the cut-off for uptake of other preventive interventions (flu vaccine (d), pneumococcal polysaccharide vaccine (PPV) (e) and statin medications (f)) as well as past shingles (b) and dementia (c) diagnoses. The data source for this analysis was the SAIL database for Wales. All analyses were run on the same sample as those for the effect of the zoster vaccine on dementia occurrence. The exception is c, for which we did not exclude individuals with a diagnosis of dementia before the start of the zoster vaccine program. The grey dots show the mean value for each 10-week increment in week of birth. The grey shading of the dots is proportionate to the weight that observations from this 10-week increment received in the analysis.

我們的分析方法主要比較那些因為在計劃開始日期前剛滿 80 歲而不符合帶狀皰疹疫苗接種資格的人，與那些因為在開始日期後剛滿 80 歲而符合接種資格的人。根據回歸不連續性分析的標準做法 ^28,29 ，實際接種疫苗的效果（與僅僅符合資格相比）是通過兩階段最小二乘回歸來確定的，該回歸將在出生日期資格門檻處結果的突然變化幅度與在此門檻處疫苗接種率的突然變化幅度進行比較（方法）。因此，並非所有符合資格的人都接種了帶狀皰疹疫苗的事實並不會對我們的分析造成偏差。

Our analysis approach primarily compares those who were ineligible for zoster vaccination because they had their 80th birthday immediately before the program’s start date with those who were eligible because they had their 80th birthday immediately after the start date. As is standard practice in regression discontinuity analyses^28,29, the effect of actually receiving the vaccine (as opposed to merely being eligible) was determined using a two-stage least-squares regression, which divides the magnitude of the abrupt change in the outcome at the date-of-birth eligibility threshold by the magnitude of the abrupt change in vaccine uptake at this threshold (Methods). Thus, the fact that not all those who were eligible received zoster vaccination does not bias our analysis.

帶狀皰疹疫苗可預防帶狀皰疹
Zoster vaccination prevents shingles

我們首先證明我們的方法成功重現了臨床試驗中已知的因果效應，即疫苗減少了帶狀皰疹的發生率 ³⁰ 。具體而言，使用回歸不連續設計 ^28,29 ，我們比較了在帶狀皰疹疫苗的出生日期資格閾值兩側出生的成年人之間帶狀皰疹的發生率。與帶狀皰疹疫苗臨床試驗中使用的方法 ³⁰ 一致，我們的結果是個體在隨訪期間是否至少有一次帶狀皰疹診斷。在我們為期 7 年的隨訪期間，我們樣本中 296,324 名成年人中共有 14,465 人至少有一次帶狀皰疹診斷。在相同的隨訪時間內，我們發現符合疫苗資格的人至少有一次帶狀皰疹診斷的概率降低了 1.0（95% CI = 0.2–1.7；P = 0.010）個百分點（圖。 2a)，相對減少 18.8%（95% CI = 8.8–28.9）。在計算實際接種帶狀皰疹疫苗的效果時，我們發現七年隨訪期間至少有一次帶狀皰疹診斷的概率減少了 2.3（95% CI = 0.5–3.9；P = 0.011）個百分點（圖 2b）；這一效果（相對而言為 37.2%（95% CI = 19.7–54.7））與活 attenuated 帶狀皰疹疫苗（Zostavax）臨床試驗中觀察到的效果相似 ³⁰ 。

We first demonstrate that our approach successfully reproduces the known causal effect from clinical trials that the vaccine reduces the occurrence of shingles³⁰. Specifically, using a regression discontinuity design^28,29, we compared the occurrence of shingles between adults born close to either side of the date-of-birth eligibility threshold for the zoster vaccine. Consistent with the approach used by clinical trials of the zoster vaccine³⁰, our outcome was whether or not an individual had at least one shingles diagnosis during the follow-up period. During our follow-up period of 7 years, a total of 14,465 among 296,324 adults in our sample had at least one diagnosis of shingles. Over the same follow-up time, we find that being eligible for the vaccine reduced the probability of having at least one shingles diagnosis by 1.0 (95% CI = 0.2–1.7; P = 0.010) percentage point (Fig. 2a), corresponding to a relative reduction of 18.8% (95% CI = 8.8–28.9). When calculating the effect of actually receiving the zoster vaccine, we find a reduction in the probability of having at least one shingles diagnosis of 2.3 (95% CI = 0.5–3.9; P = 0.011) percentage points over the seven-year follow-up period (Fig. 2b); an effect (37.2% (95% CI = 19.7–54.7) in relative terms) that is similar in size to that observed in clinical trials of the live-attenuated zoster vaccine (Zostavax)³⁰.

圖 2：帶狀皰疹疫苗對帶狀皰疹診斷的影響。

a–c，符合（a）資格和已接種（在不同的隨訪期間（b）和不同的寬限期（c））帶狀皰疹疫苗對於在隨訪期間至少有一次帶狀皰疹診斷的概率的影響估計。對於 a，MSE 最佳帶寬為 145.7 週（95,227 名成年人）。灰色點顯示每 10 週出生週的平均值。點的灰色陰影與來自這 10 週增量的觀察在分析中所獲得的權重成正比。對於 b 和 c，我們主要規範的 MSE 最佳帶寬為 116.9 週（76,316 名成年人）。三角形（而不是點）描繪了我們的主要規範。紅色（而不是白色）填充表示統計顯著性（P < 0.05）。寬限期是指自指數日期起的時間段，之後隨訪時間被認為開始。灰色垂直條顯示回歸係數點估計的 95%置信區間（雙側 t 檢驗）。

a–c, Effect estimates of being eligible for (a), and having received (across different follow-up periods (b) and across different grace periods (c)), the zoster vaccine on the probability of having at least one shingles diagnosis during the follow-up period. For a, the MSE-optimal bandwidth is 145.7 weeks (95,227 adults). The grey dots show the mean value for each 10-week increment in week of birth. The grey shading of the dots is proportionate to the weight that observations from this 10-week increment received in the analysis. For b and c, the MSE-optimal bandwidth for our primary specification is 116.9 weeks (76,316 adults). The triangles (rather than points) depict our primary specification. The red (as opposed to white) fillings denote statistical significance (P < 0.05). Grace periods refer to time periods since the index date after which the follow-up time is considered to begin. The grey vertical bars show the 95% CIs around the point estimate of the regression coefficient (two-sided t tests).

我們顯示我們的估計效果對於用來建模帶狀皰疹發生與出生週之間關係的回歸所選擇的函數形式（補充圖 5）、定義我們分析樣本的出生日期合格截止日期周圍的出生週窗口寬度（帶寬）（補充圖 6a）或不同的寬限期（圖 2c）並不敏感。所謂的「寬限期」，是指自指標日期起的時間段，之後的隨訪時間被認為開始（方法）。還有強烈的跡象表明，帶狀皰疹疫苗降低了至少一次診斷後帶狀皰疹神經痛（帶狀皰疹的常見併發症）的概率，儘管這一效果在所有規範中並未達到統計顯著性（補充圖 7）。

We show that our estimated effect is not sensitive to the chosen functional form of the regression used to model the relationship of shingles occurrence with week of birth (Supplementary Fig. 5), the width of the week-of-birth window (bandwidth) around the date-of-birth eligibility cut-off that defines our analysis sample (Supplementary Fig. 6a) or to different grace periods (Fig. 2c). With ‘grace periods’, we refer to time periods since the index date after which the follow-up time is considered to begin (Methods). There was also a strong indication that the zoster vaccine reduced the probability of having at least one diagnosis of postherpetic neuralgia (a common complication of shingles), although this effect did not reach statistical significance in all specifications (Supplementary Fig. 7).

新診斷的癡呆症
New diagnoses of dementia

考慮到不同類型癡呆症之間的神經病理重疊以及在臨床上區分癡呆症類型的困難 ³¹ ，以及在研究較不常見的結果時我們的統計能力降低，我們將癡呆症定義為任何類型或原因的癡呆症作為我們的結果。如果在我們的電子健康記錄數據中（包括所有在初級或次級護理中做出的診斷）出現新的癡呆症診斷，或者癡呆症被列為死亡證上的主要或輔助死亡原因，我們認為個體已經發展為癡呆症。在我們的七年隨訪期間，我們樣本中的 282,541 名成年人中有 35,307 人被新診斷為癡呆症。

Given the neuropathological overlap between dementia types and the difficulty in distinguishing dementia types clinically³¹, as well as our reduced statistical power when studying less-common outcomes, we defined dementia as dementia of any type or cause as our outcome. We considered an individual to have developed dementia if there was a new diagnosis of dementia in our electronic health record data (which includes all diagnoses made in primary or secondary care) or dementia was listed as a primary or contributory cause of death on the death certificate. The Read and ICD-10-codes used to define dementia are listed in the Supplementary Codes. During our seven-year follow-up period, 35,307 among 282,541 adults in our sample were newly diagnosed with dementia.

使用我們的回歸不連續性方法，我們估計符合帶狀皰疹疫苗接種資格的效果是新診斷癡呆症的概率在 7 年內絕對減少 1.3 個百分點（95% CI = 0.2–2.7；P = 0.022）和相對減少 8.5%（95% CI = 1.9–15.1）（圖 3a）。考慮到並非所有符合資格的人都接種了疫苗，我們發現實際接種帶狀皰疹疫苗使新診斷癡呆症的概率減少 3.5 個百分點（95% CI = 0.6–7.1；P = 0.019），相當於相對減少 20.0%（95% CI = 6.5–33.4）（圖 3b）。效果估計通常對不同的寬限期（圖 3c）、我們回歸的功能形式（補充圖 8）或圍繞出生日期資格截止日期的出生週窗口寬度（帶寬）（補充圖 6b）不敏感。 我們還發現帶狀皰疹疫苗對減少癡呆診斷有顯著影響，如果診斷僅定義為新開立的藥物處方（多奈哌齊鹽酸鹽、加蘭他敏、利伐斯的明或美金剛鹽酸鹽），這些藥物通常用於減緩阿茲海默症的進展（補充表 2（第 2 欄））。同樣，當調整所有輸入變數以計算癡呆風險評分 ³² （如 2013 年 9 月 1 日之前所記錄）時，效果保持相似（補充表 2（第 7 欄））。

Using our regression discontinuity approach, we estimate that the effect of being eligible for the zoster vaccine is a 1.3 (95% CI = 0.2–2.7; P = 0.022) percentage points absolute and 8.5% (95% CI = 1.9–15.1) relative reduction in the probability of a new dementia diagnosis over 7 years (Fig. 3a). Scaled to account for the fact that not all those who were eligible received the vaccine, we find that actually receiving the zoster vaccine reduced the probability of a new dementia diagnosis by 3.5 (95% CI = 0.6–7.1; P = 0.019) percentage points, corresponding to a relative reduction of 20.0% (95% CI = 6.5–33.4) (Fig. 3b). The effect estimates were generally not sensitive to different grace periods (Fig. 3c), the functional form of our regressions (Supplementary Fig. 8) nor the width of the week-of-birth window (bandwidth) drawn around the date-of-birth eligibility cut-off (Supplementary Fig. 6b). We also find significant effects of the zoster vaccine on reducing dementia diagnoses if a diagnosis is defined solely as a new prescription of a medication (donepezil hydrochloride, galantamine, rivastigmine or memantine hydrochloride) that is frequently prescribed to slow the progression of Alzheimer’s disease (Supplementary Table 2 (column 2)). Similarly, the effects remain similar when adjusting for all input variables to the Dementia Risk Score³² (as recorded before 1 September 2013) (Supplementary Table 2 (column 7)).

圖 3：帶狀皰疹疫苗對新診斷癡呆症的影響。

a–c，符合（a）資格和已接種（在不同的隨訪期間（b）和不同的寬限期（c））帶狀皰疹疫苗對新診斷癡呆症的影響估計。對於 a，MSE 最佳帶寬為 134.4 週（83,167 名成年人）。灰色點顯示每 10 週出生週的平均值。點的灰色陰影與來自這 10 週增量的觀察在分析中所獲得的權重成正比。對於 b 和 c，我們主要規範的 MSE 最佳帶寬為 90.6 週（56,098 名成年人）。三角形（而不是點）描繪了我們的主要規範。紅色（而不是白色）填充表示統計顯著性（P < 0.05）。寬限期是指自指數日期起的時間段，之後隨訪時間被認為開始。灰色垂直條顯示回歸係數點估計的 95%置信區間（雙側 t 檢驗）。

a–c, Effect estimates of being eligible for (a), and having received (across different follow-up periods (b) and across different grace periods (c)), the zoster vaccine on new diagnoses of dementia. For a, the MSE-optimal bandwidth is 134.4 weeks (83,167 adults). The grey dots show the mean value for each 10-week increment in week of birth. The grey shading of the dots is proportionate to the weight that observations from this 10-week increment received in the analysis. For b and c, the MSE-optimal bandwidth for our primary specification is 90.6 weeks (56,098 adults). The triangles (rather than points) depict our primary specification. The red (as opposed to white) fillings denote statistical significance (P < 0.05). Grace periods refer to time periods since the index date after which the follow-up time is considered to begin. The grey vertical bars show the 95% CIs around the point estimate of the regression coefficient (two-sided t tests).

其他使用相同截止點的干預措施
Other interventions using an identical cut-off

我們研究的主要優勢在於，只有當混淆變量在 1933 年 9 月 2 日的出生日期閾值 ^28,29 突然變化時，該變量才會對我們的分析產生偏見。因此，如果另一項干預也使用 1933 年 9 月 2 日的出生日期截止點作為資格標準，則可能會發生混淆偏見。這樣的干預不太可能僅影響發展癡呆的風險，而不會同時影響其他健康結果。因此，我們對 2019 年威爾士 70 歲以上年齡組的十個主要殘疾調整生命年和死亡原因，實施了與我們對帶狀皰疹和癡呆所採用的相同回歸不連續性方法 ³³ ，以及所有屬於查爾森共病指數的條件 ³⁴ 。如補充圖 9 和 10 所示，我們通常未檢測到帶狀皰疹疫苗對這些其他常見健康結果的新診斷的影響。

The key strength of our study is that a confounding variable can bias our analysis only if the variable changes abruptly at the 2 September 1933 date-of-birth threshold^28,29. Thus, confounding bias could occur if another intervention also used the date of birth cut-off of 2 September 1933 as an eligibility criterion. Such an intervention is unlikely to affect only the risk of developing dementia without also influencing other health outcomes. We therefore implemented the same regression discontinuity approach as we did for shingles and dementia for the ten leading causes of disability-adjusted life years and mortality for the age group 70+ years in Wales in 2019³³, and all conditions that are part of the Charlson Comorbidity Index³⁴. As shown in Supplementary Figs. 9 and 10, we generally do not detect effects of zoster vaccination on new diagnoses of these other common health outcomes.

我們進行了四種額外的分析，所有這些分析都提供了證據，表明另一項干預措施使用了與帶狀皰疹疫苗推出的出生日期合格門檻相同的日-月-年組合（1933 年 9 月 2 日）。首先，我們顯示 1933 年 9 月 2 日的出生日期門檻不影響接受其他預防健康干預措施的概率（補充圖 11）。其次，我們檢查了用於帶狀皰疹疫苗合格的日-月（即 1933 年 9 月 2 日）出生日期截止是否也被其他影響癡呆風險的干預措施使用。我們通過對 2013 年 9 月 1 日（帶狀皰疹疫苗計劃實際開始的日期）進行相同的分析，來實現這一點，並對 2013 年前後的三年中的每一年的 9 月 1 日進行比較。因此，例如，當將計劃的開始日期移至 2012 年 9 月 1 日時，我們將 1932 年 9 月 2 日出生日期門檻附近的人與從 2012 年 9 月 1 日開始的隨訪期間進行比較。 作為額外的檢查，使我們能夠維持在主要分析中使用的七年隨訪期的長度，我們將計劃開始日期移至 2013 年前的六年中每年的 9 月 1 日。正如預期的那樣，對於這兩項檢查，我們發現只有在實際由帶狀皰疹疫苗接種計劃使用的出生日期截止點（1933 年 9 月 2 日）上，對癡呆症發生有顯著影響（補充圖 12 和 13）。第三，我們發現，在帶狀皰疹疫苗推出前的七年期間，1933 年 9 月 2 日出生日期閾值周圍的年齡群體之間，癡呆症的七年發病率沒有差異（補充圖 14）。第四，使用 2011 年人口普查的數據，我們在補充圖 15-17 中顯示，在威爾士達到特定教育水平的個體比例在 1933 年 9 月 2 日的閾值上沒有不連續性。

We undertook four additional types of analysis, all of which provide evidence against another intervention having used the identical day-month-year combination (2 September 1933) as was used as the date-of-birth eligibility threshold for the zoster vaccine rollout. First, we show that the 2 September 1933 date-of-birth threshold does not affect the probability of taking up other preventive health interventions (Supplementary Fig. 11). Second, we examined whether the day–month (that is, 2 September) date-of-birth cut-off used for zoster vaccine eligibility was also used by other interventions that affect dementia risk. We did so by implementing the identical analysis as for 1 September 2013 (the actual date on which the zoster vaccine program started) for 1 September of each of the three years before and after 2013. Thus, for example, when shifting the start date of the program to 1 September 2012, we compared those around the 2 September 1932 date-of-birth threshold with the follow-up period starting on 1 September 2012. As an additional check that enabled us to maintain the length of the seven-year follow-up period used in our primary analyses, we shifted the program start date to 1 September of each of the 6 years preceding (but not after) 2013. As expected, for both of these checks, we find a significant effect on dementia occurrence only for the date-of-birth cutoff (2 September 1933) that was actually used by the zoster vaccination program (Supplementary Figs. 12 and 13). Third, we find that there is no difference in the seven-year incidence of dementia between age cohorts around the 2 September 1933 date-of-birth threshold for the seven-year period before the zoster vaccine rollout (Supplementary Fig. 14). Fourth, using data from the 2011 Census, we show in Supplementary Figs. 15–17 that there are no discontinuities across the 2 September 1933 threshold in the proportion of individuals in Wales who reached a particular level of education.

對不同分析方法的穩健性
Robustness to a different analytical approach

作為我們研究結果穩健性的額外測試，我們實施了所有主要分析，使用了差異中的差異工具變量分析（DID-IV），這利用了我們預期結果會在 1933 年 9 月 2 日的出生日期閾值上發生突變的事實（即，帶狀皰疹疫苗接種計劃所使用的資格截止日期的日–月–年組合）。這樣做的過程中，我們的分析放寬了回歸不連續性的連續性假設（即，潛在混淆變量在 1933 年 9 月 2 日的出生資格閾值上不會顯示出突然變化的假設），而是假設（在沒有帶狀皰疹疫苗接種計劃的情況下）在 1933 年 9 月 2 日的閾值上結果的可能不連續性與之前出生年份的 9 月 2 日閾值上的不連續性在大小上沒有差異。我們的方法詳情在方法部分中提供。 令人鼓舞的是，在我們七年的追蹤期間，帶狀皰疹疫苗接種對新診斷癡呆症的概率的影響在 DID-IV 和回歸不連續性方法之間非常相似（−3.1（95% CI = −5.8 至−0.4，P = 0.024）對比−3.5（95% CI = −7.1 至−0.6，P = 0.019）個百分點）（圖 4）。帶狀皰疹和後帶狀皰疹神經痛的結果也是如此（圖 4）。我們對 DID-IV 的比較組之間的健康特徵平衡進行了與回歸不連續性分析相同的檢查（補充圖 18）。我們還驗證了我們的 DID-IV 方法僅對癡呆症、帶狀皰疹和後帶狀皰疹神經痛的結果產生顯著影響，而對其他常見健康結果則沒有（補充圖 18）。

As an additional test of the robustness of our findings, we implemented all primary analyses using a difference-in-differences instrumental variable analysis (DID-IV) that takes advantage of the fact that the only 2 September date-of-birth threshold at which we would expect an abrupt change in the outcome is the 2 September threshold in 1933 (that is, the day–month–year combination that was used as eligibility cut-off by the zoster vaccination program). In doing so, our analysis relaxes the continuity assumption of regression discontinuity (that is, the assumption that potential confounding variables do not display a sudden change at the 2 September 1933 date-of-birth eligibility threshold), and instead assumes that (in the absence of the zoster vaccination program) a possible discontinuity in the outcome at the 2 September 1933 threshold is not different in size from a discontinuity at the 2 September threshold in previous birth years. Details of our approach are provided in the Methods. Encouragingly, the effect of zoster vaccine receipt on the probability of a new dementia diagnosis during our seven-year follow-up period is remarkably similar between the DID-IV and regression discontinuity approach (−3.1 (95% CI = −5.8 to −0.4, P = 0.024) versus −3.5 (95% CI = −7.1 to −0.6, P = 0.019) percentage points) (Fig. 4). This is also the case for the outcomes of shingles and postherpetic neuralgia (Fig. 4). We conducted the same checks for balance in health characteristics between our comparison groups for the DID-IV as we implemented for our regression discontinuity analyses (Supplementary Fig. 18). We also verified that our DID-IV approach yields significant effects only for the outcomes of dementia, shingles and postherpetic neuralgia, but not for any other common health outcomes (Supplementary Fig. 18).

圖 4：DID-IV 與回歸不連續方法之間效果估計的比較。

比較接受帶狀皰疹疫苗對新診斷癡呆症、帶狀皰疹和後帶狀皰疹神經痛的絕對效應估計，基於 DID-IV 和回歸不連續性分析。此分析的數據來源為威爾士的 SAIL 數據庫。癡呆症結果的樣本量為 96,767 名成年人，而帶狀皰疹和後帶狀皰疹神經痛結果的樣本量為 105,258 名成年人。P 值是使用雙側 t 檢驗計算的。DID-IV 對帶狀皰疹的效應的 P 值為 0.001。誤差條表示回歸係數的點估計周圍的 95%置信區間（雙側 t 檢驗）。

Comparison of absolute effect estimates of having received the zoster vaccine on new diagnoses of dementia, shingles and postherpetic neuralgia between the DID-IV and the regression discontinuity analyses. The data source for this analysis was the SAIL database for Wales. The sample size for the dementia outcome is 96,767 adults and the sample for the shingles and postherpetic neuralgia outcomes is 105,258 adults. P values were calculated using two-sided t-tests. The P value for the DID-IV effect on shingles is 0.001. The error bars depict the 95% CIs around the point estimate of the regression coefficient (two-sided t-tests).

效應機制的探索
Explorations of the effect mechanism

帶狀皰疹疫苗對癡呆症診斷的保護效應可能源於三種（非互斥）機制：（1）由於帶狀皰疹事件而導致的醫療途徑變化；（2）水痘帶狀皰疹病毒（VZV）再激活的減少；以及（3）VZV 無關的免疫調節效應（例如，通過異源適應性免疫或訓練的先天免疫介導的效應）。在本節中，我們提供證據來檢驗這些機制。

A protective effect of zoster vaccination on dementia diagnoses could arise from three (non-mutually exclusive) mechanisms: (1) changes in healthcare pathways as a result of a shingles episode; (2) a reduction in reactivations of the varicella zoster virus (VZV); and (3) a VZV-independent immunomodulatory effect (for example, one mediated through heterologous adaptive immunity or trained innate immunity). In this section, we present evidence to examine each of these mechanisms.

由於接種帶狀皰疹疫苗而減少的醫療使用，可能導致健康系統有較少的機會去 (1) 診斷癡呆症（確定偏差）；或 (2) 實施護理變更（例如，開始使用新藥物），這會增加未來被診斷為癡呆症的風險。重要的是要指出，這種機制不太可能完全解釋我們的發現，因為我們對接種帶狀皰疹疫苗所減少的帶狀皰疹發作的效應估計值相當小，無法合理解釋觀察到的癡呆症診斷減少。

Reduced healthcare use resulting averted shingles episodes from zoster vaccination receipt could have translated to fewer opportunities for the health system to (1) diagnose dementia (ascertainment bias); or (2) implement care changes (for example, initiation of a new medication) that increase the risk of being diagnosed with dementia in the future. It is important to point out that this mechanism is unlikely to fully explain our findings, because the size of our effect estimates for reductions in shingles episodes from zoster vaccination were considerably too small to plausibly account for the observed reduction in dementia diagnoses.

我們 nonetheless 進行了五種類型的分析，以進一步檢查這一潛在機制。首先，如果帶狀皰疹發作為健康系統提供了診斷癡呆的機會，那麼它們也可能提供診斷其他慢性病的機會。因此，我們將與帶狀皰疹和癡呆相同的回歸不連續性方法應用於 2019 年威爾士 70 歲以上年齡組中，十個主要導致殘疾調整生命年和死亡率的慢性病，或查爾森共病指數的一部分。我們在隨訪期間以一年為增量繪製了我們的估計。除了風濕病外，我們顯示符合帶狀皰疹疫苗接種資格並未對新慢性病診斷產生影響（補充圖 19）。第二，我們在隨訪期間調整了回歸模型，以考慮健康服務使用的頻率（初級護理就診次數、門診就診次數、住院次數和接種流感疫苗的次數），這並未實質性改變我們的效應估計（補充表 2（第 4 列））。 第三，我們在將分析隊列限制為 247,784 名（佔我們主要分析隊列的 87.6%）在帶狀皰疹疫苗推出前的 5 年內每年至少拜訪一次其初級保健提供者的患者時實施了我們的分析。這項分析的理由是，在那些已經經常與健康系統互動的患者中，由於避免帶狀皰疹發作而減少的一次與健康系統的接觸不太可能影響檢測未診斷癡呆症的概率。這一頻繁使用醫療保健的隊列中的效應大小與我們的主要分析隊列中的效應大小相似（補充表 2（第 3 列））。第四，我們在主要回歸不連續性分析中添加了個體在隨訪期間是否經歷帶狀皰疹發作作為協變量。我們發現，調整我們的分析以考慮帶狀皰疹發作並未實質性改變我們的點估計（補充圖 20）。 第五，我們在主要回歸不連續性分析中，對於在隨訪期間接受帶狀皰疹診斷的參與者，實施了一項事件研究，這些參與者位於均方誤差（MSE）最佳帶寬內。為了調查帶狀皰疹的發作是否導致患者所接受的醫療服務發生變化，我們檢查了帶狀皰疹診斷對診斷後 36 個月內以下結果的影響：(1) 接受新癡呆診斷的概率；(2) 一組健康服務使用的指標；(3) 接受抗病毒藥物、鴉片類藥物、加巴噴丁或普瑞巴林的新藥物處方的概率，以及在 SAIL 數據庫的另一項分析中與癡呆風險增加相關的 216 種藥物中的任何一種；(4) 被診斷為任何屬於查爾森共病指數的慢性病的概率。 我們發現，帶狀皰疹診斷並未增加在帶狀皰疹診斷後幾個月內獲得新癡呆診斷的概率，並且僅導致醫療服務使用和新藥物處方的短期增加（補充圖 21）。在帶狀皰疹發作後幾個月內獲得加巴噴丁或普瑞巴林處方的概率雖然持續增加，但幅度較小。同樣，帶狀皰疹發作當月與發作前一月相比，診斷任何慢性病的概率增加不到一個百分點（補充圖 21）。

We nonetheless conducted five types of analysis to examine this potential mechanism further. First, if shingles episodes presented an opportunity for the health system to diagnose dementia, then they would probably also present an opportunity to diagnose other chronic conditions. We therefore applied the same regression discontinuity approach as for shingles and dementia to all chronic conditions that are either among the ten leading causes of disability-adjusted life years and mortality for the age group 70+ years in Wales in 2019³³ or part of the Charlson Comorbidity Index³⁴. We plotted our estimates across one-year increments in the follow-up period. With the exception of rheumatological diseases, we show that being eligible for the zoster vaccine did not have an effect on new chronic disease diagnoses (Supplementary Fig. 19). Second, we adjusted our regressions for the frequency of health service use (the number of primary care visits, outpatient visits, hospital admissions and influenza vaccinations received) during the follow-up period, which did not substantially change our effect estimates (Supplementary Table 2 (column 4)). Third, we implemented our analyses when restricting the analysis cohort to the 247,784 (87.6% of the analysis cohort for our primary analyses) patients who visited their primary care provider at least once a year during each of the 5 years before the start of the zoster vaccine rollout. The rationale for this analysis is that, among patients who already interact frequently with the health system, a reduction of one further contact with the health system due to an averted shingles episode is less likely to affect the probability of detecting undiagnosed dementia. The effect sizes among this cohort of frequent healthcare users remain similar to those in our primary analytical cohort (Supplementary Table 2 (column 3)). Fourth, we added whether individuals experienced a shingles episode during the follow-up period as a covariate in our primary regression discontinuity analysis. We found that adjusting our analysis for shingles episodes did not substantially change our point estimate (Supplementary Fig. 20). Fifth, we implemented an event study among those participants in the mean-squared-error (MSE)-optimal bandwidth of our primary regression discontinuity analysis for dementia who received a shingles diagnosis during the follow-up period. To investigate whether episodes of shingles led to changes in healthcare received by patients, we examined the effect of the shingles diagnosis on the following outcomes in each of the 36 months after the diagnosis: (1) the probability of receiving a new dementia diagnosis; (2) a set of indicators of health service use; (3) the probability of receiving a new medication prescription for antiviral drugs, opioid medications, gabapentin or pregabalin, and any of 216 medications that were associated with an increased risk of dementia in another analysis in the SAIL database²³; and (4) the probability of being diagnosed with any of the chronic conditions that are part of the Charlson Comorbidity Index³⁴. We found that shingles diagnoses did not increase the probability of receiving a new dementia diagnosis in the months after the shingles diagnosis, and led to only short-term increases in healthcare service use and new medication prescriptions (Supplementary Fig. 21). The increase in the probability of receiving a gabapentin or pregabalin prescription in the months after the shingles episode, while more sustained, was small in magnitude. Similarly, the increase in the probability of being diagnosed with any chronic condition in the month of a shingles episode compared with the month before the episode was less than one percentage point (Supplementary Fig. 21).

由於帶狀皰疹疫苗對帶狀皰疹發作的影響是中等的（圖 2），本節中的五種類型分析僅記錄了帶狀皰疹發作對醫療途徑的小且短暫的影響，即使對這些護理途徑對癡呆症影響的最保守假設也暗示，帶狀皰疹發作所導致的醫療變化無法解釋我們的發現。

As the effect of zoster vaccination on shingles episodes is moderate (Fig. 2), and the five types of analysis in this section document only small and short-lived effects of shingles episodes on healthcare pathways, even the most conservative assumptions about the effect of these care paths on dementia imply that changes in healthcare as a result of a shingles episode cannot explain our findings.

如前一節所述，調整我們的回歸不連續性分析以考慮患者在隨訪期間是否有至少一次帶狀疱疹發作的記錄並未實質性改變我們的點估計（補充圖 20）。然而，這項分析關於水痘帶狀疱疹病毒再激活減少作為效應機制的結論受到以下事實的限制：（1）帶狀疱疹疫苗接種可能減少水痘帶狀疱疹病毒的臨床和亞臨床再激活；（2）帶狀疱疹發作可能不是整個隨訪期間經歷的亞臨床水痘帶狀疱疹病毒再激活程度的可靠指標，因為帶狀疱疹發作可能會增強水痘帶狀疱疹病毒的免疫力。我們因此進行了以下分析，以進一步檢查水痘帶狀疱疹病毒再激活減少作為效應機制。

As described in the previous section, adjusting our regression discontinuity analysis for whether a patient had a record of at least one shingles episode during the follow-up period did not change our point estimate substantially (Supplementary Fig. 20). However, conclusions from this analysis regarding reductions in VZV reactivations as the effect mechanism are limited by the fact that (1) zoster vaccination probably reduces both clinical as well as subclinical reactivations of VZV^30,35; and (2) having a shingles episode may be an unreliable indicator of the degree of subclinical VZV reactivations experienced during the entire follow-up period, given that shingles episodes may boost immunity for VZV^30,35. We therefore conducted the following analyses to further examine reductions in VZV reactivations as the effect mechanism.

首先，我們檢查了在隨訪期間，帶狀皰疹疫苗對癡呆症影響開始的時間。具體而言，在出生日期接近 1933 年 9 月 2 日的患者中，我們繪製了符合與不符合帶狀皰疹疫苗接種資格的癡呆症的 Kaplan–Meier 和累積發生率曲線（方法）。如果影響機制是通過減少水痘帶狀皰疹病毒（VZV）的再激活，那麼人們會預期疫苗對臨床和亞臨床病毒再激活的減少效果會在觀察到對癡呆症的影響之前開始。活疫苗帶狀皰疹疫苗被認為在疫苗接種後幾週內開始有效 ^30,36 。與 VZV 再激活的影響應該先於癡呆症影響的原則一致，我們觀察到癡呆症發生率的減少僅在一年以上後才開始出現，無論是在整體人群中還是在僅女性中（補充圖 22）。

First, we examined the time during the follow-up period at which the effect of zoster vaccination on dementia appears to begin. Specifically, among patients who were born in close proximity to the 2 September 1933 date-of-birth threshold, we plotted the Kaplan–Meier and cumulative incidence curves for dementia for those who were eligible versus ineligible for zoster vaccination (Methods). If the effect mechanism is through a reduction in VZV reactivations, then one would expect that the effects of the vaccine on reductions in clinical and subclinical reactivations of the virus would begin before observing an effect on dementia. The live-attenuated zoster vaccine is thought to begin being efficacious within weeks after vaccine administration^30,36. Consistent with the principle that the effect on VZV reactivations should precede the dementia effect, we observe that the reduction in the incidence of dementia begins to emerge only after more than one year, both among the full population as well as among women only (Supplementary Fig. 22).

其次，雖然帶狀皰疹發作可能增強水痘帶狀疱疹病毒（VZV）免疫力，因此減少隨後的亞臨床 VZV 再活化 ^30,35 ，但在隨訪期間經歷多次發作的個體與僅經歷一次帶狀皰疹發作的個體相比，可能在隨訪期間經歷更高程度的臨床和亞臨床 VZV 再活化 ³⁰ 。因此，我們使用傾向得分匹配（方法）比較了經歷多次與一次帶狀皰疹發作對癡呆症的關聯。我們發現經歷多次帶狀皰疹發作的人癡呆症的發病率更高（補充圖 23）。

Second, while a shingles episode may boost VZV immunity and, therefore, reduce subsequent subclinical VZV reactivations^30,35, individuals who experience multiple episodes as opposed to a single shingles episode during the follow-up period probably experience a greater degree of both clinical and subclinical VZV reactivations during the follow-up period³⁰. Using propensity score matching (Methods), we therefore compared the association with dementia from experiencing multiple versus a single shingles episode. We find a higher incidence of dementia among those who experienced multiple shingles episodes (Supplementary Fig. 23).

第三，如果水痘帶狀疱疹病毒的再激活增加了癡呆的風險，那麼通過抗病毒藥物限制病毒在帶狀疱疹發作期間的複製程度，預期可以減少癡呆的發生率。因此，我們使用多變量 Cox 比例風險模型（方法），比較了接受抗病毒藥物治療的帶狀疱疹患者與未接受治療的患者之間與癡呆的關聯。我們發現，帶狀疱疹發作的抗病毒治療與癡呆的發生率降低相關（補充圖 23）。

Third, if VZV reactivations increase the risk of dementia, then limiting the degree of replication of the virus during a shingles episode through antiviral medication could be expected to decrease dementia incidence. Using a multivariable Cox proportional hazards model (Methods), we therefore compared the association with dementia between individuals whose shingles episode was treated with antiviral medication and those for whom the episode was untreated. We find that antiviral treatment of a shingles episode is associated with a reduced incidence of dementia (Supplementary Fig. 23).

為了探究這一機制，我們利用文獻中關於疫苗接種的病原體無關免疫調節效應的兩個觀察結果：它們往往（1）在性別上有很大差異，活疫苗接種的有益效果通常僅在女性中出現，而男性則不然 ^6,7,8 ；（2）取決於在接種該疫苗之前或與之同時接種其他疫苗 ^6,7,8 。與這些觀察結果一致，我們發現帶狀皰疹疫苗對新診斷癡呆症的影響在女性中明顯大於男性（圖 5 和補充表 3（第 1 列））。在帶狀皰疹疫苗對帶狀皰疹和後帶狀皰疹神經痛的診斷影響上，女性和男性之間沒有顯著差異（補充表 3（第 2 和第 3 列））。同樣，在 1933 年 9 月 2 日的出生日期合格門檻上，疫苗接種的突然增加幅度在女性和男性之間是相當的（補充圖 24），男性的幅度略大。

To probe this mechanism, we take advantage of two observations on pathogen-independent immunomodulatory effects from vaccination in the literature: they tend to (1) vary strongly by sex, with beneficial effects from live-attenuated vaccination often seen only in female but not male individuals^6,7,8; and (2) depend on the receipt of other vaccines before, or at the same time as, receipt of the vaccine in question^6,7,8. Consistent with these observations, we find that the effect of zoster vaccination on new diagnoses of dementia was markedly greater among women than men (Fig. 5 and Supplementary Table 3 (column 1)). There was no significant difference between women and men in the effect of the zoster vaccine on diagnoses of shingles and postherpetic neuralgia (Supplementary Table 3 (columns 2 and 3)). Similarly, the magnitude of the abrupt increase in vaccine uptake at the 2 September 1933 date-of-birth eligibility threshold was comparable between women and men (Supplementary Fig. 24), with a slightly larger magnitude among men.

圖 5：帶狀皰疹疫苗對女性和男性新診斷癡呆症的影響。

討論 Discussion

在這裡我們發現，帶狀皰疹疫苗在七年的隨訪期間將新診斷癡呆症的概率降低了約五分之一。通過利用帶狀皰疹疫苗在威爾士推出的獨特方式構成了一個自然實驗的事實，並檢查每一個可能的偏見來源，我們的研究提供了比現有的僅僅是關聯性的證據更有可能是因果性的證據。 我們的顯著效應大小，加上帶狀皰疹疫苗相對較低的成本，意味著如果這些發現確實是因果性的，那麼帶狀皰疹疫苗在預防或延遲癡呆症方面將比現有的藥物干預更有效且更具成本效益。

Here we found that the zoster vaccine reduced the probability of a new dementia diagnosis by approximately one-fifth over a seven-year follow-up period. By taking advantage of the fact that the unique way in which the zoster vaccine was rolled out in Wales constitutes a natural experiment, and examining each possible remaining source of bias, our study provides evidence that is more likely to be causal in nature than the existing, exclusively associational^{15,16,17,18,19,20,21,22,23,24}, evidence on this topic. Our substantial effect sizes, combined with the relatively low cost of the zoster vaccine, imply that, if these findings are truly causal, the zoster vaccine will be both far more effective as well as cost-effective in preventing or delaying dementia than existing pharmaceutical interventions.

我們的準實驗方法減少了與更標準的關聯分析相比混淆的可能性。此外，我們提供了一系列分析的證據，反對任何可能的偏見來源成為我們發現的可能解釋。儘管如此，我們的發現仍有可能（即使在統計上不太可能）是由於偶然。因而，在其他人群、環境和數據來源中確認我們的發現至關重要。重要的是，我們成功地使用英格蘭和威爾士的全國死亡證數據確認了我們的發現 ¹⁰ 。具體而言，由於英格蘭以幾乎與威爾士相同的方式推出帶狀皰疹疫苗 ⁴⁴ ，我們能夠使用與我們來自威爾士的電子健康記錄數據相同的準實驗方法，來確定根據出生日期的帶狀皰疹疫苗接種資格對於記錄為癡呆的死亡的影響。我們發現，在九年的隨訪期間，大約每 20 例此類死亡中就有 1 例是因為符合帶狀皰疹疫苗接種資格而避免的。 這項研究對我們的結果構成了重要的確認，因為它分析了不同的人群（英格蘭的人口約佔英格蘭和威爾士總人口的 95% ⁴⁵ ）、數據類型（死亡證明而非電子健康記錄）和結果（因癡呆症死亡）。除了對我們的死亡數據結果的確認外，偶然發現的可能性進一步降低，因為我們成功地使用第二種分析方法（DID-IV）重複了我們的主要發現，並且我們的效應大小在多種分析選擇中保持穩定，包括寬限期的選擇、隨訪期、研究人群定義（例如，限制為經常使用醫療服務的人）、回歸的功能形式、圍繞出生日期合格截止日期的出生週窗口的寬度以及指數日期定義。

Our quasi-experimental approach reduces the probability of confounding compared with more standard associational analyses. Moreover, we have provided evidence from a series of analyses against any of the possible remaining sources of bias being a likely explanation of our findings. Nonetheless, it is possible (even if statistically unlikely) that our findings are due to chance. Confirmation of our findings in other populations, settings and data sources is therefore critical. Importantly, we have successfully confirmed our findings using country-wide death certificate data from England and Wales¹⁰. Specifically, because England rolled out the zoster vaccine in an almost identical way to Wales⁴⁴, we were able to use the same quasi-experimental approach as in our electronic health record data from Wales to determine the effect of eligibility for zoster vaccination based on one’s date of birth on deaths for which the underlying cause was recorded as being dementia. We found that, over a nine-year follow-up period, approximately 1 in 20 such deaths were averted from being eligible for zoster vaccination. This study constitutes an important confirmation of our results because it analysed a different population (England’s population accounts for approximately 95% of England’s and Wales’s combined population⁴⁵), type of data (death certificates as opposed to electronic health records) and outcome (deaths due to dementia). In addition to this confirmation of our results in mortality data, the probability of a chance finding is further reduced by the fact that we successfully replicate our main findings using a second analysis approach (DID-IV) and that our effect sizes remain stable across a multitude of analysis choices, including choice of grace periods, follow-up periods, study population definitions (for example, restriction to frequent healthcare users), functional form of our regressions, width of the week-of-birth window drawn around the date-of-birth eligibility cut-off and index date definitions.

我們觀察到帶狀皰疹疫苗對女性和男性在癡呆症影響上的巨大差異，女性的受益程度高於男性。在我們看來，這些女性和男性之間的巨大差異有幾個合理的原因。首先，我們不能排除男性因接種帶狀皰疹疫苗而導致新癡呆症診斷大幅減少的可能性，特別是考慮到我們的數據中男性在老年時癡呆症的發病率低於女性，因此我們對男性的分析信心區間更寬。其次，疫苗的非靶向效應通常在女性中比男性更強，特別是女性在活減毒疫苗中受益更多 ^6,7 。第三，對疫苗的免疫反應似乎存在重要的性別差異 ⁴⁶ 。最後，越來越多的證據表明，女性和男性在癡呆症的發病機制上可能存在差異 ⁴⁷ 。

We observed large differences in the effect of zoster vaccination on dementia between women and men, with women benefitting more than men. In our view, these large differences between women and men are plausible for several reasons. First, we cannot exclude the possibility of substantial reductions in new dementia diagnoses from zoster vaccination among men, especially given the lower incidence of dementia in older age among men than women in our data and, therefore, our wider confidence intervals for analyses among men. Second, off-target effects of vaccines have often been observed to be far stronger among female than male individuals, with female individuals benefiting more from live-attenuated vaccines in particular^6,7. Third, there appear to be important sex differences in the immunological response to vaccines more generally⁴⁶. Lastly, there is a growing body of evidence that there may be differences in the pathogenesis of dementia between women and men⁴⁷.

除了投資隨機試驗外，對水痘帶狀疱疹病毒（VZV）及其對帶狀疱疹疫苗的免疫反應在癡呆病理學中潛在作用的基礎科學研究的投資，可能提供關鍵的機制性見解。已有幾條證據顯示，VZV 的再激活與癡呆之間存在合理的機制性途徑。具體而言，已發現 VZV 的再激活會通過血管病變 ^48,49 、淀粉樣蛋白沉積和 tau 蛋白聚集 ⁵⁰ 、神經炎症 ^51,52,53,54 ，以及與阿茲海默症相似的腦血管疾病譜，包括小到大血管疾病、缺血、梗塞和出血 ^{51,52,53,54,55,56} ，導致持久的認知障礙。正如最近的一項研究所建議的 ⁵⁷ ，減少 VZV 的亞臨床和臨床再激活可能會通過神經炎症途徑減少腦中單純疱疹病毒-1 的再激活。這一機制將 VZV 與有關單純疱疹病毒-1 在癡呆病理學中作用的文獻相聯繫 ^1,2,3,4,5 。 儘管如此，我們對帶狀皰疹疫苗與癡呆之間聯繫的效應機制的探索性分析表明，通過減少水痘帶狀皰疹病毒的臨床和亞臨床再激活的機制以及一種與病原體無關的免疫機制都是合理的。這些可能的與病原體無關的免疫機制中的一些最近在其他地方有詳細說明 ⁵⁸ 。

Other than investing into randomized trials, investments into basic science research on the potential role of VZV and the immune response to the zoster vaccine in the pathogenesis of dementia could provide critical mechanistic insights. There are already several lines of evidence on plausible mechanistic pathways that link VZV reactivations to dementia. Specifically, VZV reactivations have been found to lead to long-lasting cognitive impairment through vasculopathy^48,49, amyloid deposition and aggregation of tau proteins⁵⁰, neuroinflammation^51,52,53,54, as well as a similar spectrum of cerebrovascular disease as seen in Alzheimer’s disease, including small to large vessel disease, ischaemia, infarction and haemorrhage^{51,52,53,54,55,56}. As suggested by a recent study⁵⁷, it may also be the case that reducing subclinical and clinical reactivations of VZV reduces reactivations of the herpes simplex virus-1 in the brain through neuroinflammatory pathways. This mechanism would link VZV to the body of literature on the role of herpes simplex virus-1 in the pathogenesis of dementia^1,2,3,4,5. Nonetheless, our exploratory analyses on the effect mechanism that links zoster vaccination to dementia suggest that both a mechanism through reducing clinical and subclinical reactivations of VZV as well as a pathogen-independent immune mechanism are plausible. Some of these possible pathogen-independent immune mechanisms have recently been detailed elsewhere⁵⁸.

我們的研究有幾個限制。首先，我們的結果確認可能存在某種程度的漏檢，無論是在癡呆症的診斷是否以及多及時的情況下。重要的是，由於漏檢癡呆症的概率以及延遲診斷的情況不太可能在 1933 年 9 月 2 日的帶狀皰疹疫苗接種資格門檻上突然改變，因此這一結果的錯誤分類最有可能是非差異性的。因此，我們的效應估計可能是保守的（也就是說，我們的絕對效應大小將低估真實的絕對效應幅度）。同樣，在我們的隨訪期間，癡呆症確認的準確性和及時性隨著年份的變化而變化，例如由於臨床實踐或健康系統激勵措施的變化，影響了 1933 年 9 月 2 日之前和之後出生的人。因此，我們不會期望這些變化成為我們分析中的偏差來源。 其次，我們無法提供帶狀皰疹疫苗在減少其他年齡組癡呆症發生率方面的有效性估計，這些年齡組在我們的回歸不連續性分析中權重最重（主要是 79 至 80 歲的人群）。第三，COVID-19 大流行可能影響了癡呆症的診斷及時性。然而，我們主要分析中使用的隨訪期間在 COVID-19 大流行開始之前結束。此外，由於大流行對 1933 年 9 月 2 日之前和之後出生的人影響相同，因此大流行相關的癡呆症檢測不足不太可能對我們的相對效應估計產生偏見。第四，我們的隨訪期最多限制為 8 年。因此，我們的研究無法提供有關帶狀皰疹疫苗在此時間段之外減少癡呆症發生率的有效性的信息。最後，由於較新的重組亞單位帶狀皰疹疫苗（Shingrix）在英國僅於 2023 年 9 月取代了活減毒帶狀皰疹疫苗（Zostavax），而這是在我們的隨訪期結束之後，因此我們的效應估計僅適用於活減毒帶狀皰疹疫苗。

Our study has several limitations. First, our outcome ascertainment probably suffers from some degree of under-detection, both in whether and in how timely a fashion dementia is diagnosed. Importantly, because the probability of under-detecting dementia, as well as the delay in doing so, is unlikely to change abruptly at the 2 September 1933 date-of-birth eligibility threshold for zoster vaccination, this outcome misclassification is most likely non-differential. Our effect estimates are therefore likely to be conservative (that is, our absolute effect sizes would be an underestimate of the true absolute effect magnitude). Similarly, changes in the accuracy and timeliness of dementia ascertainment over the years of our follow-period, such as due to changing clinical practice or health system incentives to detect and record dementia, affected those born immediately before versus immediately after 2 September 1933 equally. We would therefore not expect these changes to be a source of bias in our analyses. Second, we are unable to provide estimates for the effectiveness of the zoster vaccine for reducing dementia occurrence in age groups other than those who were weighted most heavily in our regression discontinuity analyses (primarily those aged 79 to 80 years). Third, the COVID-19 pandemic probably affected the timeliness with which dementia was diagnosed. However, the follow-up period used in our primary analyses ended before the start of the COVID-19 pandemic. Moreover, because the pandemic affected those born just before versus just after 2 September 1933 equally, pandemic-related under-detection of dementia is unlikely to bias our relative effect estimates. Fourth, we were limited to a maximum follow-up period of 8 years. Our study can therefore not inform on the effectiveness of the zoster vaccine for reducing dementia occurrence beyond this time period. Lastly, because the newer recombinant subunit zoster vaccine (Shingrix) replaced the live-attenuated zoster vaccine (Zostavax) in the United Kingdom only in September 2023²⁵, which is after our follow-up period ended, our effect estimates apply to the live-attenuated zoster vaccine only.

方法 Methods

活疫苗帶狀皰疹疫苗（Zostavax）於 2013 年 9 月 1 日起通過分階段推廣系統向符合條件的個體提供。在此系統下，年滿 71 歲或以上的個體每年 9 月 1 日被劃分為三組：（1）不符合資格的群體，年齡在 71 至 78 歲（或根據計劃年份的不同為 77 歲），未來將符合資格；（2）補救群體，包含年齡為 79 歲（或 78 歲，根據計劃年份的不同）；（3）因年齡在 80 歲或以上而不符合資格的個體，且從未符合資格。

The live-attenuated zoster vaccine (Zostavax) was made available to eligible individuals in Wales through a staggered rollout system starting on 1 September 2013. Under this system, individuals aged 71 years or older were categorized into three groups on 1 September of each year: (1) an ineligible cohort of those aged 71 to 78 years (or 77 years, depending on the year of the program), who became eligible in the future; (2) a catch-up cohort, consisting of individuals aged 79 years (or 78 years, again depending on the year of the program); and (3) those who were ineligible as they were aged 80 years or older and who never became eligible.

我們的分析集中在 1925 年 9 月 1 日（計劃開始時 88 歲）至 1942 年 9 月 1 日（計劃開始時 71 歲）之間出生的成年人。1925 年 9 月 1 日至 1933 年 9 月 1 日之間出生的人從未符合資格，而 1933 年 9 月 2 日至 1942 年 9 月 1 日之間出生的人則在補救隊列中逐漸符合資格。具體而言，疫苗在計劃的第一年（2013 年 9 月 1 日至 2014 年 8 月 31 日）提供給 1933 年 9 月 2 日至 1934 年 9 月 1 日之間出生的人；在第二年（2014 年 9 月 1 日至 2015 年 8 月 31 日）提供給 1934 年 9 月 2 日至 1936 年 9 月 1 日之間出生的人；在第三年（2015 年 9 月 1 日至 2016 年 8 月 31 日）提供給 1936 年 9 月 2 日至 1937 年 9 月 1 日之間出生的人；在第四年（2016 年 9 月 1 日至 2017 年 8 月 31 日）提供給 1937 年 9 月 2 日至 1938 年 9 月 1 日之間出生的人。截至 2017 年 4 月 1 日，個體在其 78 歲生日時符合疫苗接種資格，並在其 80 歲生日之前保持符合資格。我們的分析主要比較了在 1933 年 9 月 2 日或稍後出生的個體與那些在 1933 年 9 月 2 日之前出生而從未符合資格的個體。 我們在補充圖 26 至 28 中顯示，大多數符合資格的個體，特別是在我們分析的前兩個資格群體中，在他們符合資格的第一年內接種了疫苗（而不是在後來的年份），而這兩個資格群體的疫苗接種率相似。

Our analysis focused on adults born between 1 September 1925 (88 years old at program start) and 1 September 1942 (71 years old at program start). Those born between 1 September 1925 and 1 September 1933 never became eligible, whereas those born between 2 September 1933 and 1 September 1942 became progressively eligible in a catch-up cohort. Specifically, the vaccine was offered to those born between 2 September 1933 and 1 September 1934 in the first year of the program (1 September 2013 to 31 August 2014); those born between 2 September 1934 and 1 September 1936 in the second year (1 September 2014 to 31 August 2015); those born between 2 September 1936 and 1 September 1937 in the third year (1 September 2015 to 31 August 2016); and those born between 2 September 1937 and 1 September 1938 in the fourth year (1 September 2016 to 31 August 2017). As of 1 April 2017, individuals become eligible for the vaccine on their 78th birthday and remain eligible until their 80th birthday. Our analysis principally compared individuals born on or shortly after 2 September 1933, to individuals who never became eligible as they were born shortly before 2 September 1933. We show in Supplementary Figs. 26–28 that most eligible individuals, especially in the first two eligibility cohorts, which are the focus of our analysis, took up the vaccination during their first year of eligibility (as opposed to during later years) and that vaccination uptake in these first two eligibility cohorts was of a similar magnitude.

威爾士的醫療服務是通過威爾士國民健康服務（NHS）提供的，該服務是英國單一支付者單一提供者醫療系統的一部分 ⁵⁹ 。NHS 威爾士和斯旺西大學創建了 SAIL 數據庫 ^{26,60,61,62,63,64} ，該數據庫包括與醫院護理信息以及該國死亡登記數據相關聯的初級護理訪問的完整電子健康記錄數據。

Healthcare in Wales is provided through the Welsh National Health Service (NHS), which is part of the United Kingdom’s single-payer single-provider healthcare system⁵⁹. NHS Wales and Swansea University created the SAIL Databank^{26,60,61,62,63,64}, which includes full electronic health record data for primary care visits linked to information on hospital-based care as well as the country’s death register data.

SAIL 生成了一個所有曾經在威爾士註冊的初級保健提供者的個體名單（這對於居住在威爾士的成年人來說超過 98% ²⁷ ），來自威爾士人口服務數據集 ⁶⁵ 。然後，SAIL 將這些個體與以下每個數據集進行連結。來自初級保健提供者的電子健康記錄數據通過威爾士縱向全科醫療數據集 ⁶⁶ 提供，該數據集包含來自約 80% 的威爾士初級保健診所和 83% 的威爾士人口的數據。這些電子健康記錄數據使用 Read 代碼，提供有關患者及其護理接觸的詳細信息，包括診斷、臨床徵象和觀察、症狀、實驗室檢查和結果、執行的程序以及行政項目 ⁶⁷ 。帶狀皰疹疫苗接種的定義使用了疫苗接種的管理代碼和產品代碼（補充表 1）。 在醫院環境中進行的診斷和程序（作為住院入院或日間手術的一部分）通過與威爾士患者事件數據庫的連接在 SAIL 中提供，該數據庫始於 1991 年，包含威爾士所有醫院基礎護理的數據，以及在英格蘭提供給威爾士居民的醫院基礎護理。程序使用 OPCS-4 代碼編碼，診斷使用 ICD-10 代碼編碼。任何 NHS 威爾士醫院門診部的出席信息通過與威爾士門診數據庫的連接提供，該數據庫始於 2004 年。癌症的 ICD-10 編碼診斷通過與威爾士癌症情報和監測單位的連接來識別，該單位是威爾士的國家癌症登記處，記錄所有提供給威爾士居民的癌症診斷，無論他們在哪裡被診斷或治療。該數據集始於 1994 年。最後，通過與年度地區死亡提取的連接，為所有威爾士居民提供死亡原因數據（無論他們在英國的哪裡去世），該數據始於 1996 年，並包括死亡證明中的主要和輔助死亡原因。 死亡日期是指登記死亡的日期，而不是實際發生的日期。死亡原因數據在 2001 年之前使用 ICD-9 編碼，之後使用 ICD-10 編碼。

SAIL generates a list of all individuals who have ever been registered with a primary care provider in Wales (which is the case for over 98% of adults residing in Wales²⁷) from the Welsh Demographic Service Dataset⁶⁵. SAIL then links this universe of individuals to each of the following datasets. Electronic health record data from primary care providers is made available in SAIL through the Welsh Longitudinal General Practice dataset⁶⁶, which contains data from approximately 80% of primary care practices in Wales and 83% of the Welsh population. These electronic health record data use Read codes, which provide detailed information on patients and their care encounters, including diagnoses, clinical signs and observations, symptoms, laboratory tests and results, procedures performed and administrative items⁶⁷. Zoster vaccination was defined using both codes for the administration of the vaccine as well as product codes (Supplementary Table 1). Diagnoses made and procedures performed in the hospital setting (as part of inpatient admissions or day-case procedures) are provided in SAIL through linkage to the Patient Episode Database for Wales⁶⁸, which begins in 1991 and contains data for all hospital-based care in Wales as well as hospital-based care provided in England to Welsh residents. Procedures are encoded using OPCS-4 codes⁶⁹ and diagnoses using ICD-10 codes⁷⁰. Attendance information at any NHS Wales hospital outpatient department is provided through linkage to the Outpatient Database for Wales⁷¹, which starts in 2004. ICD-10-encoded diagnoses of cancers are identified through linkage to the Welsh Cancer Intelligence and Surveillance Unit⁷², which is the national cancer registry for Wales that records all cancer diagnoses provided to Welsh residents wherever they were diagnosed or treated. This dataset begins in 1994. Finally, cause-of-death data are provided for all Welsh residents (regardless of where they died in the United Kingdom) through linkage to the Annual District Death Extract⁷³, which begins in 1996 and includes primary and contributory causes of death from death certificates. Dates for deaths were those on which the death was registered, as opposed to when it occurred. Cause-of-death data use ICD-9 coding until 2001 and ICD-10 coding thereafter.

在測試教育成就是否存在任何不連續性時，我們使用了由國家統計局（ONS）提供的數據集 ⁷⁴ 。該數據集是由 ONS 根據 2011 年英國人口普查生成的，涵蓋所有年滿 16 歲的常住居民，這些居民在 1925 年 1 月至 1950 年 12 月之間出生於威爾士，無論其就業狀況如何。ONS 根據性別、出生月份和年份（1925 年 1 月至 1950 年 12 月）、最高學歷和職業對數據進行了分類。

When testing for any discontinuities in educational attainment across the date-of-birth eligibility threshold, we used a dataset provided by the Office for National Statistics (ONS)⁷⁴. This dataset was generated by the ONS from the 2011 UK Census for all usual residents aged 16 or over, born in Wales between January 1925 and December 1950, regardless of their employment status. The data were categorized by the ONS by sex, month and year of birth (January 1925 to December 1950), highest level of qualification and occupation.

信息治理審查小組（IGRP，申請編號 1306）已批准倫理審查。IGRP 由政府、監管機構和專業機構組成，負責監督和批准使用 SAIL 數據庫的申請。所有分析均已獲得斯坦福大學機構審查委員會的批准，並於 2023 年 6 月 9 日被認為風險最小（協議編號 70277）。

Ethics approval was granted by the Information Governance Review Panel (IGRP, application number, 1306). Composed of government, regulatory and professional agencies, the IGRP oversees and approves applications to use the SAIL databank. All analyses were approved and considered minimal risk by the Stanford University Institutional Review Board on 9 June 2023 (protocol number, 70277).

我們的研究人群由 296,603 名在 1925 年 9 月 1 日至 1942 年 9 月 1 日之間出生的個體組成，他們在帶狀皰疹疫苗計劃啟動日（2013 年 9 月 1 日）時已在威爾士的初級保健提供者處登記。由於我們僅能獲得個體出生周的星期一的日期，因此無法確定在 1933 年 8 月 28 日開始的截止周出生的個體是否在疫苗計劃啟動的第一年內符合接種帶狀皰疹疫苗的資格。因此，我們排除了 279 名在這一特定周出生的個體。在剩餘的個體中，有 13,783 名在 2013 年 9 月 1 日之前已被診斷為癡呆，因此被排除在以新診斷癡呆為結果的分析之外。因此，我們最終的分析隊列大小為 282,541。這一分析隊列用於所有以新癡呆診斷為結果的主要分析，除了那些以疫苗計劃啟動前的癡呆發病率、帶狀皰疹和後帶狀皰疹神經痛為結果的分析；對於這些分析，我們並未排除在 2013 年 9 月 1 日之前已診斷為癡呆的個體。

Our study population consisted of 296,603 individuals born between 1 September 1925 and 1 September 1942 who were registered with a primary care provider in Wales on the start date of the zoster vaccine program rollout (1 September 2013). As we only had access to the date of the Monday of the week in which an individual was born, we were unable to determine whether the individuals born in the cut-off week starting on 28 August 1933 were eligible for the zoster vaccine in the first year of its rollout. We therefore excluded 279 individuals born in this particular week. Among the remaining individuals, 13,783 had a diagnosis of dementia before 1 September 2013 and were therefore excluded from the analyses with new diagnoses of dementia as outcome. The size of our final analysis cohort for all primary analyses for new dementia diagnoses was therefore 282,541. This analysis cohort was used for all analyses except those with incidence of dementia before zoster vaccination program start, shingles and postherpetic neuralgia as outcomes; analyses for which we did not exclude individuals with a dementia diagnosis before 1 September 2013.

我們跟踪了從 2013 年 9 月 1 日到 2021 年 8 月 31 日的個體，這使得最大跟踪期為 8 年。在我們的主要規範中，我們選擇了 7 年的跟踪期（即直到 2020 年 8 月 31 日），因為這使我們能夠包括長達 12 個月的寬限期，同時保持在出生日期合格截止日期兩側的個體的跟踪期不變。然而，我們也顯示了從 1 年到 8 年的跟踪期的所有結果，並以一年為增量。由於每位患者都被分配了獨特的匿名 NHS 號碼，我們能夠隨著時間跟踪個體，即使他們更換了初級護理提供者。因此，只有當患者移民出威爾士或更換到約 20%不向 SAIL 提供數據的威爾士初級護理診所時，我們的隊列才會失去跟踪。在我們的七年跟踪期內，這種情況發生在我們主要分析隊列中的 23,049 名（8.2%）成年人中，且在 1933 年 9 月 2 日合格門檻之前和之後出生的個體之間，這一比例沒有顯著差異。 在我們的主要分析隊列中，總共有 92,629 名（37.8%）成年人在七年的隨訪期間去世。我們的主要分析方法不考慮任何死亡的競爭風險，原因有三。首先，在沒有帶狀皰疹疫苗接種計劃的情況下，死亡的競爭風險不應在 1933 年 9 月 2 日的出生日期合格門檻之間有所不同。其次，在我們的情境中不調整死亡的競爭風險是一個保守的選擇，因為帶狀皰疹疫苗接種的合格性可能會降低（但不太可能增加）全因死亡率 ^10,75 。因此，符合帶狀皰疹疫苗接種資格的人，平均來說，將會暴露於更長的時間段，期間他們可能會被新診斷為癡呆症。第三，迄今為止，在回歸不連續框架中，尚不存在成熟的生存分析方法，包括確定 CACE 和最佳帶寬的能力 ⁷⁶ 。

We followed individuals from 1 September 2013 to 31 August 2021, which allowed for a maximum follow-up period of 8 years. In our primary specification, we selected a follow-up period of 7 years (that is, until 31 August 2020) because this enabled us to include grace periods of up to 12 months while still keeping the follow-up period constant for individuals on either side of the date-of-birth eligibility cut-off. However, we also show all results for follow-up periods from one to eight years in one-year increments. Owing to the unique anonymized NHS number assigned to each patient, we were able to follow individuals across time even if they changed primary care provider. Patients were therefore only lost to follow-up in our cohort if they emigrated out of Wales or changed to one of the approximately 20% of primary care practices in Wales that did not contribute data to SAIL. Over our seven-year follow-up period, this was the case for 23,049 (8.2%) of adults in our primary analysis cohort, with no significant difference in this proportion between those born just before versus just after the 2 September 1933 eligibility threshold. In total, 92,629 (37.8%) of adults in our primary analysis cohort died during the seven-year follow-up period. Our primary analysis approach does not adjust for any competing risk of death for three reasons. First and foremost, in the absence of a zoster vaccination program, there is no reason that the competing risk of death should differ across the 2 September 1933 date-of-birth eligibility threshold. Second, not adjusting for competing risk of death in our setting is a conservative choice because eligibility for zoster vaccination may reduce (but is very unlikely to increase) all-cause mortality^10,75. Thus, those eligible for zoster vaccination will, on average, be exposed to a longer time period during which they could become newly diagnosed with dementia. Third, to date, no well-established approach exists for survival analysis in a regression discontinuity framework, including the ability to determine the CACE and optimal bandwidth⁷⁶.

由於不同類型癡呆症之間的神經病理重疊以及在臨床上區分癡呆症類型的困難，我們選擇將癡呆症定義為任何類型或原因的癡呆症。癡呆症被定義為在初級護理（如初級護理電子健康記錄數據中所記錄）、專科護理或醫院護理中作出的癡呆症診斷，或在死亡證明上被列為主要或輔助死亡原因的癡呆症。首次在這些數據來源中記錄癡呆症的日期用於定義患者被診斷為癡呆症的日期。同樣，所有其他結果也使用在任何護理環境中作出的診斷或提及作為主要或輔助死亡原因來定義。對於慢性病，首次在這些數據來源中記錄的日期用於定義慢性病發生的日期。 對於非慢性病症或事件（即帶狀皰疹、後帶狀皰疹神經痛、中風、下呼吸道感染、跌倒、下背痛、藥物處方、流感疫苗接種和乳腺癌篩查），在隨訪期間，使用程序日期之後在任何這些數據來源中首次記錄的日期來定義結果的發生。用於定義所有結果的 Read 和 ICD-10 代碼詳情見補充代碼。

Owing to the neuropathological overlap between dementia types and difficulty in distinguishing dementia types clinically^77,78,79, we chose to define dementia as dementia of any type or cause. Dementia was defined as a diagnosis of dementia made either in primary care (as recorded in the primary care electronic health record data), specialist care or hospital-based care, or dementia being named as a primary or contributory cause of death on the death certificate. The date of the first recording of dementia across any of these data sources was used to define the date on which the patient was diagnosed with dementia. Similarly, all other outcomes were defined using a diagnosis made in any care setting or mentioning as a primary or contributory cause of death. For chronic conditions, the date of the first recording across any of these data sources was used to define the date on which the chronic condition occurred. For non-chronic conditions or events (that is, shingles, postherpetic neuralgia, stroke, lower respiratory tract infections, falls, lower back pain, medication prescriptions, influenza vaccination and breast cancer screening), the date of first recording after the program date across any of these data sources was used for defining the occurrence of the outcome during the follow-up. The Read and ICD-10 codes used to define all outcomes are detailed in the Supplementary Codes.

兩位分析數據的作者（M.E.和 M.X.）獨立編碼了分析的所有部分。由於不同的數據編碼選擇，偶爾會出現小的差異，這些差異通過討論得到了解決。

The two authors who analysed the data (M.E. and M.X.) have coded all parts of the analysis independently. Occasional minor differences, resulting from different data coding choices, were resolved through discussion.

我們的回歸不連續性方法
Our regression discontinuity approach

我們使用回歸不連續設計來分析我們的數據，這是一種在社會科學中用於因果推斷的成熟方法 ⁸⁰ 。回歸不連續分析估計在截止點左側和右側的預期結果概率，以獲得治療效果的估計。我們在主要分析中使用了局部線性三角核回歸（對靠近出生日期合格閾值的觀察賦予更高的權重），並在穩健性檢查中使用了二次多項式。這是回歸不連續分析的推薦和最穩健的方法，即使在分配變量（此處為出生日期）與結果之間的關係是指數型的情況下也是如此 ⁸¹ 。回歸不連續分析中的一個重要選擇是圍繞閾值繪製的數據窗口的寬度（帶寬）。根據標準做法，我們在主要分析中使用了 MSE 最優帶寬 ⁸² ，這最小化了回歸擬合的均方誤差。我們為每個樣本和結果定義的組合單獨確定了這個最優帶寬。 在穩健性檢查中，我們檢查了我們的點估計在不同帶寬選擇下的變化程度，這些帶寬範圍從 0.25 倍到兩倍的 MSE 最佳帶寬。我們使用了穩健的偏差修正標準誤差進行推斷 ⁸³ 。

We used a regression discontinuity design to analyse our data, which is a well-established method for causal inference in the social sciences⁸⁰. Regression discontinuity analysis estimates expected outcome probabilities just left and just right of the cut-off, to obtain an estimate of the treatment effect. We used local linear triangular kernel regressions (assigning a higher weight to observations lying closer to the date-of-birth eligibility threshold) in our primary analyses and quadratic polynomials in robustness checks. This is the recommended and most robust approach for regression discontinuity analyses even in situations in which the relationship between the assignment variable (here, date of birth) and the outcome is exponential⁸¹. An important choice in regression discontinuity analyses is the width of the data window (the bandwidth) that is drawn around the threshold. Following standard practice, we used an MSE-optimal bandwidth⁸², which minimizes the MSEs of the regression fit, in our primary analyses. We determined this optimal bandwidth separately for each combination of sample and outcome definition. In robustness checks, we examined the degree to which our point estimates vary across different bandwidth choices ranging from 0.25 times to two times the MSE-optimal bandwidth. We used robust bias-corrected standard errors for inference⁸³.

估算符合帶狀皰疹疫苗接種資格的影響
Estimating the effect of being eligible for the zoster vaccine

在第一步中，我們確定了有資格接種帶狀皰疹疫苗（無論個人是否實際接種疫苗）對我們結果的影響。為此，我們估計了以下回歸方程：
In the first step, we determined the effect of being eligible for the zoster vaccine (regardless of whether the individual actually received the vaccine) on our outcomes. To do so, we estimated the following regression equation:

其中 Y 是一個二元變數，若個體經歷了結果（例如，帶狀皰疹或癡呆症）則等於一。二元變數 D 表示是否符合帶狀皰疹疫苗的接種資格，若個體出生於 1933 年 9 月 2 日或之後則等於一。術語 (WOB − C ₀ ) 表示個體的出生週，圍繞著截止日期。交互項 D × (WOB − C ₀ ) 允許回歸線的斜率在閾值的兩側有所不同。參數 β ₁ 確定符合疫苗接種資格對結果的絕對影響。無論我們報告相對影響，我們都是通過將絕對影響估計值 β ₁ 除以出生日期資格閾值左側的平均結果來計算的，即 α 的估計值。

where Y_i is a binary variable equal to one if an individual experienced the outcome (for example, shingles or dementia). The binary variable D_i indicates eligibility for the zoster vaccine and is equal to one if an individual was born on or after the cut-off date of 2 September 1933. The term (WOB_i − C₀) indicates an individual’s week of birth centred around the cut-off date. The interaction term D_i × (WOB_i − C₀) allows for the slope of the regression line to differ on either side of the threshold. The parameter β₁ identifies the absolute effect of being eligible for the vaccine on the outcome. Wherever we report relative effects, we calculated these by dividing the absolute effect estimate β₁ by the mean outcome just left of the date-of-birth eligibility threshold, that is, the estimate of α.

估計實際接種帶狀皰疹疫苗的效果
Estimating the effect of actually receiving the zoster vaccine

在第二步中，我們估計了實際接種帶狀皰疹疫苗對我們結果的影響。這種影響在計量經濟學文獻中通常被稱為合規者平均因果效應（CACE）。按照標準做法，我們使用模糊回歸不連續設計來估計 CACE。模糊回歸不連續分析考慮到疫苗並不是在出生周的截止日期上確定性分配的。相反，一部分不符合資格的個體仍然接種了疫苗，而一部分符合資格的個體則沒有接種疫苗。為了考慮這種分配的模糊性，模糊回歸不連續設計使用了工具變量方法，工具變量是指示個體是否有資格接種疫苗的二元變量，即在 1933 年 9 月 2 日或之後出生。正如我們在出生周的疫苗接種圖中驗證的那樣（圖。 1a)，在出生日期資格門檻之後出生的個體接種帶狀皰疹疫苗的概率遠高於在門檻之前出生的個體。除了接種帶狀皰疹疫苗的概率的突然變化之外，出生在資格門檻之前和之後的個體之間可能沒有其他影響我們結果發生概率的特徵差異。因此，出生日期資格門檻的指標變量是一個有效的工具變量，用於識別接種帶狀皰疹疫苗對我們結果的因果影響。為了比較實際接種帶狀皰疹疫苗的個體與未接種者之間經歷結果的概率，工具變量估計通過在出生日期資格門檻處接種疫苗的概率的突然變化來縮放符合接種帶狀皰疹疫苗資格的效應大小。跳躍的大小通過以下第一階段回歸方程估計：

In the second step, we estimated the effect of actually receiving the zoster vaccine on our outcomes. This effect is commonly referred to as the complier average causal effect (CACE) in the econometrics literature⁸⁴. As is standard practice⁸⁴, we used a fuzzy regression discontinuity design to estimate the CACE. Fuzzy regression discontinuity analysis takes into account the fact that the vaccine is not deterministically assigned at the week-of-birth cut-off. Instead, a proportion of ineligible individuals still received the vaccine and a proportion of eligible individuals did not receive the vaccine. To account for this fuzziness in the assignment, the fuzzy regression discontinuity design uses an instrumental variable approach, with the instrumental variable being the binary variable that indicates whether or not an individual was eligible to receive the vaccine, that is, is born on or after 2 September 1933. As we verify in our plot of vaccine receipt by week of birth (Fig. 1a), individuals who were born immediately after the date-of-birth eligibility threshold had a far higher probability of receiving the zoster vaccine compared with those born immediately before the threshold. Other than the abrupt change in the probability of receiving the zoster vaccine, there probably is no other difference in characteristics that affect the probability of our outcomes occurring between those born immediately after versus immediately before the date-of-birth eligibility threshold. Thus, the indicator variable for the date-of-birth eligibility threshold is a valid instrumental variable to identify the causal effect of receipt of the zoster vaccine on our outcomes. To compare the probability of experiencing the outcome between those who actually received the zoster vaccine versus those who did not, the instrumental variable estimation scales the effect size for being eligible for the zoster vaccine by the size of the abrupt change in the probability of receiving the vaccine at the date-of-birth eligibility threshold. The size of the jump is estimated through the following first-stage regression equation:

其中 V 是一個二元變數，表示個體是否接種了帶狀皰疹疫苗，而 θ ₁ 標識在出生日期資格閾值處疫苗接種的非連續增長。所有其他參數與回歸（1）中的相同。
where V_i is a binary variable indicating whether the individual received the zoster vaccine and θ₁ identifies the discontinuous increase in vaccine receipt at the date-of-birth eligibility threshold. All other parameters are the same as in regression (1).

通過將資格的效果與方程（2）中的第一階段效果重新縮放來估計的 CACE 可以表示為對 μ ₁ 的 IV 估計，來自以下第二階段回歸：
The CACE estimated by rescaling the effect of eligibility with the first-stage effect from equation (2) can be represented as an IV estimate for μ₁ from the following second-stage regression:

其中預測帶狀皰疹疫苗接種的概率是從方程式()的第一階段估計中獲得的。這個 CACE，代表了在遵從者中接種疫苗的（絕對）平均因果效應，即只有在符合條件的情況下才會接種疫苗的患者。
其中 ({hat{V}}_{i}) 是從方程 (2) 的第一階段估計中獲得的帶狀皰疹疫苗接種的預測概率。這個 CACE，μ ₁ ，代表在遵從者中接種疫苗的（絕對）平均因果效應，即只有在符合條件的情況下才接種疫苗的患者。
其中 是從方程 () 的第一階段估計中獲得的帶狀皰疹疫苗接種的預測概率。這個 CACE，代表了在遵從者中接種疫苗的（絕對）平均因果效應，即只有在符合條件的情況下才接種疫苗的患者。
where V^i is the predicted probability of zoster vaccine receipt obtained from the first-stage estimation from equation (2). This CACE, μ₁, represents the (absolute) average causal effect of receiving the vaccine among compliers, that is, patients who take up the vaccine if and only if they are eligible.

為了計算相對效應大小，我們首先引入一些符號。設未接種疫苗的遵從者的平均結果為，接種疫苗的遵從者在臨界點的平均結果為。根據定義，絕對 CACE 為=−，相對效應為 。為了估計相對效應，我們需要對進行估計。雖然無法單獨識別遵從者，但我們可以估計他們可觀察特徵的均值，包括（參考）。設所有接種疫苗的個體（包括遵從者）在臨界點的平均結果為。假設不存在違反者（只有在不符合資格的情況下才接種疫苗的患者），所有接種疫苗的人要麼是遵從者，要麼是始終接種者（不論其資格如何都接種疫苗的患者）。因此，等於接種疫苗的遵從者和始終接種者的平均結果的加權平均：=×+×，其中和是遵從者和始終接種者的人口比例，則是在臨界點的始終接種者的平均結果。解出得 。此方程式中所有右側的數量均可從數據中估計。 首先，可以分別估算為 + 和，僅在接種疫苗的個體中重新估算回歸 ()。其次，可以分別估算為 和 ，來自回歸 ()。最後，我們通過上述公式估算 = R − 。相對效應估算為 。這些步驟中涉及的所有回歸可以堆疊並聯合估算，因此相對效應表達為已知估計量的可微函數，使用 delta 方法可以估算相對 CACE 的 95% 置信區間。
為了計算相對效應大小，我們首先引入一些符號。令 R _0,c 為未接種疫苗的遵從者的平均結果，R _1,c 為剛好在閾值處接種疫苗的遵從者的平均結果。根據定義，絕對 CACE 為 μ ₁ = R _1,c − R _0,c ，相對效應為 (frac{{mu }__1,c}{{R}_{0,c}})。要估計相對效應，我們需要 R _0,c 的估計值。雖然無法單獨識別遵從者，但我們可以估計他們可觀察特徵的均值，包括 R _0,c （參考 ⁸⁵ ）。令 R ₁ 表示在截止點所有接種疫苗的個體（包括遵從者）的平均結果。假設不存在違反者（只有在不符合資格的情況下才接種疫苗的患者），所有接種疫苗的人要麼是遵從者，要麼是始終接種者（不論其資格如何都接種疫苗的患者）。因此，R ₁ 等於接種疫苗的遵從者和始終接種者的平均結果的加權平均：R ₁ = P _c × R _1 ,c + P _a × R _1 ,a ，其中 P _c 和 P _a 是遵從者和始終接種者的人口比例，R _1 ,a 是截止點始終接種者的平均結果。 解出 R _1 ,c 得到 ({R}_{1,c}=frac{{R}__1,c-{R}_{1,a}times {P}_{a}}{{P}_{c},})。這個方程式右側的所有量都可以從數據中估算出來。首先，R ₁ 和 R _1 ,a 可以分別估算為 α + β ₁ 和 α，僅在接種疫苗的個體中重新估算回歸 (1)。其次，P _a 和 P _c 可以分別估算為 (frac{gamma }{{theta }__1,c+gamma }) 和 (frac{{theta }__1,c}{{theta }__1,c+gamma }) 來自回歸 (2)。最後，我們通過上述公式估算 R _1 ,c 和 R _0,c = R _1,c − μ ₁ 。相對效應估算為 (frac{{mu }__1,c}{{R}_{0,c}})。這些步驟中涉及的所有回歸可以堆疊並聯合估算，因此相對效應表達為已知估計量的可微函數，使用 delta 方法可以估算相對 CACE 的 95% 置信區間 ⁸⁶ 。
為了計算相對效應大小，我們首先引入一些符號。設未接種疫苗的遵從者的平均結果為，接種疫苗的遵從者在閾值處的平均結果為。根據定義，絕對 CACE 為=−，相對效應為。為了估計相對效應，我們需要對進行估計。雖然無法單獨識別遵從者，但我們可以估計他們可觀察特徵的均值，包括（參考）。設所有接種疫苗的個體（包括遵從者）在截止點的平均結果為。假設不存在違反者（只有在不符合資格的情況下才接種疫苗的患者），所有接種疫苗的人要麼是遵從者，要麼是始終接種者（不論其資格如何都接種疫苗的患者）。因此，等於接種疫苗的遵從者和始終接種者的平均結果的加權平均：=×+×，其中和是遵從者和始終接種者的人口比例，則是截止點上始終接種者的平均結果。解出得。此方程式中所有右側的數量均可從數據中估計。 首先，可以分別估計為 + 和，僅在接種疫苗的個體中重新估計回歸 ()。其次，可以分別估計為 和，來自回歸 ()。最後，我們通過上述公式估計 = R − 。相對效應被估計為 。所有涉及這些步驟的回歸可以堆疊並共同估計，因此相對效應表達為已知估計量的可微函數，使用 delta 方法可以估計相對 CACE 的 95% 置信區間。

To compute relative effect sizes, we first introduce some notation. Let R_0,c be the mean outcome among unvaccinated compliers and R_1,c the mean outcome among vaccinated compliers just at the threshold. By definition, the absolute CACE is μ₁ = R_1,c − R_0,c and the relative effect is μ1R0,c. To estimate the relative effect, we need an estimate for R_0,c. While it is impossible to identify compliers individually, we can estimate means of their observable characteristics, including R_0,c (ref. ⁸⁵). Let R₁ denote the mean outcome among all vaccinated individuals (including compliers) at the cut-off. Assuming no defiers exist (patients who get vaccinated if and only if they are not eligible), all vaccinated people are either compliers or always-takers (patients who get vaccinated irrespective of their eligibility). Thus, R₁ is equal to the population-weighted average of the mean outcomes among vaccinated compliers and always-takers: R₁ = P_c × R_1,c + P_a × R_1,a, where P_c and P_a are the population share of the compliers and always-takers and R_1,a is the mean outcome among always-takers at the cut-off. Solving for R_1,c yields R1,c=R1−R1,a×PaPc. All right-hand-side quantities in this equation can be estimated from data. First, R₁ and R_1,a can be estimated, respectively, as α + β₁ and α from re-estimating regression (1) only among vaccinated individuals. Second, P_a and P_c can be estimated, respectively, as γθ1+γ and θ1θ1+γ from regression (2). Finally, we estimate R_1,c by the above formula and R_0,c = R_1,c − μ₁. The relative effect is estimated as μ1R0,c. All regressions involved in these steps can be stacked and jointly estimated, so that the relative effect is expressed as a differentiable function of known estimators a 95% confidence interval of the relative CACE can be estimated using the delta method⁸⁶.

分析以調查另一項干預是否使用相同的出生日期資格截止日期
Analyses to investigate whether another intervention used the identical date-of-birth eligibility cut-off

我們的分析只有在混淆變數在 1933 年 9 月 2 日的出生日期資格門檻突然改變的情況下才會受到混淆，以至於非常接近這一門檻兩側的個體將不再可以互換。最可能的這種混淆變數的情境是存在一項干預，該干預使用與帶狀皰疹疫苗推出完全相同的出生日期資格門檻，並且在我們的隨訪期間也影響了癡呆診斷的概率。我們進行了五項分析，以證明這種干預的存在不太可能，通過確立受此類干預影響的結果和行為的測量在出生日期資格截止點上是平滑的。

Our analysis can only be confounded if the confounding variable changes abruptly at the 2 September 1933 date-of-birth eligibility threshold such that individuals very close to either side of this threshold would no longer be exchangeable with each other. The most plausible scenario of such a confounding variable would be the existence of an intervention that used the exact same date-of-birth eligibility threshold as the zoster vaccine rollout and that also affected the probability of a dementia diagnosis during our follow-up period. We conducted five analyses to demonstrate that the existence of such an intervention is unlikely, by establishing that measures of outcomes and behaviours that would be affected by such an intervention are smooth across the date-of-birth eligibility cut-off.

首先，在 1933 年 9 月 2 日的資格門檻附近的多個出生日期範圍內，我們繪製了在帶狀皰疹疫苗計劃開始之前（2013 年 9 月 1 日）接受以下診斷或干預的概率：帶狀皰疹診斷、在前 12 個月內接種流感疫苗、作為成年人接種肺炎球菌疫苗、目前使用他汀類藥物（定義為在計劃開始前 3 個月內的新處方或重複處方他汀類藥物）、目前使用抗高血壓藥物（定義為在計劃開始前 3 個月內的新處方或重複處方抗高血壓藥物）、參加乳腺癌篩查（定義為有轉診記錄、參加乳腺癌篩查或乳腺攝影報告的女性比例）、2019 年威爾士的十個主要殘疾調整生命年和死亡原因，這些數據由全球疾病負擔項目估算 ³³ ，以及所有合併症（艾滋病除外，因為它屬於 SAIL 數據庫未提供的隱私保護診斷）這些合併症包含在廣泛使用的查爾森合併症指數中 ³⁴ 。 此外，我們使用了這些條件中的每一個，包括性別、威爾士多重剝奪指數的十分位數 ⁵⁶ ，以及所有輸入變數來計算癡呆風險評分（如 2013 年 9 月 1 日之前所記錄的） ³² ，以預測在我們的主要回歸不連續性分析中，對於每位患者在 MSE 最佳帶寬下新癡呆診斷的概率（同時假設固定年齡）。除了繪製這些預測概率在 1933 年 9 月 2 日資格門檻附近的各種出生日期的分佈外，我們還繪製了符合與不符合帶狀皰疹疫苗接種資格的患者的這些預測概率的分佈。這些變數的 Read 代碼在補充表 1 和表 2 中提供。與臨床試驗中的平衡表一樣，這些圖表提供了保證，表明接近 1933 年 9 月 2 日資格門檻的個體彼此之間可能是可互換的。

First, across a range of birthdates around the 2 September 1933 eligibility threshold, we plotted the probability of having received the following diagnoses or interventions before the start of the zoster vaccine program (on 1 September 2013): diagnosis of shingles, influenza vaccine receipt in the preceding 12 months, receipt of the pneumococcal vaccine as an adult, current statin use (defined as a new or repeat prescription of a statin in the 3 months preceding program start), current use of an antihypertensive medication (defined as a new or repeat prescription of an antihypertensive drug in the 3 months preceding the program start), participation in breast cancer screening (defined as the proportion of women with a record of referral to, attendance at or a report from breast cancer screening or mammography), each of the ten leading causes of disability-adjusted life years and mortality for Wales in 2019 as estimated by the Global Burden of Disease Project³³, and all comorbidities (except for AIDS, which falls under privacy-protected diagnoses not made available by the SAIL database) that are included in the widely-used Charlson Comorbidity Index³⁴. Moreover, we used each of these conditions, gender, decile of Welsh Index of Multiple Deprivation⁵⁶, as well as all input variables to the Dementia Risk Score (as recorded before 1 September 2013)³², to predict the probability (while imputing a fixed age) of a new dementia diagnosis for each patient in the MSE-optimal bandwidth in our primary regression discontinuity analysis for dementia. In addition to plotting these predicted probabilities across a range of birthdates around the 2 September 1933 eligibility threshold, we also plotted the distribution of these predicted probabilities for patients who were eligible versus patients who were ineligible for zoster vaccination. The Read codes for each of these variables are provided in Supplementary Tables 1 and 2. As is the case for balance tables in clinical trials, these plots provide reassurance that individuals close to either side of the 2 September 1933 eligibility threshold are likely to be exchangeable with each other.

其次，我們對於在 1933 年 9 月 2 日門檻兩側生日的個體進行了相同的分析，還對於在 1933 年前後三年中生日接近 9 月 2 日的人進行了分析。例如，當將計劃的開始日期移至 2011 年 9 月 1 日時，我們在 2011 年 9 月 1 日開始了隨訪期，並比較了在 1931 年 9 月 2 日的資格門檻附近的個體。為了確保這些比較中的隨訪長度相同，我們不得不將這組分析的隨訪期縮短至 5 年。因此，作為額外的檢查，我們將計劃的開始日期移至 2013 年前六年的每年 9 月 1 日（但不包括 2013 年），這使我們能夠保持與主要分析相同的七年隨訪期。如果另一項影響癡呆風險的干預措施也使用 9 月 2 日的門檻來定義資格，那麼我們可能會期望在這些比較中觀察到對其他出生年份的 9 月 2 日門檻附近個體的癡呆發病率的影響。

Second, we conducted the same analysis as we did for individuals with birthdays on either side of the 2 September 1933 threshold also for people with birthdays around 2 September of each of the three years of birth preceding and succeeding 1933. For example, when moving the start date of the program to 1 September 2011, we started the follow-up period on 1 September 2011 and compared individuals around the 2 September 1931 eligibility threshold. To ensure the same length of follow-up in each of these comparisons, we had to reduce the follow-up period to 5 years for this set of analyses. Thus, as an additional check, we shifted the start date of the program to 1 September of each of the six years preceding (but not succeeding) 2013, which enabled us to maintain the same seven-year follow-up period as in our primary analysis. If another intervention that affects dementia risk also used the 2 September threshold to define eligibility, we may then expect to observe effects on dementia incidence for these comparisons of individuals just around the 2 September thresholds of other birth years.

第三，我們進行了與我們主要分析中相同的比較，針對 1933 年 9 月 2 日的出生日期門檻，除了將隨訪期間的開始時間提前到帶狀皰疹疫苗計劃推出前的 7 年。如果有一項在帶狀皰疹疫苗計劃推出之前就已經實施的干預措施，並且使用了 1933 年 9 月 2 日的出生日期合格門檻，那麼我們可以預期在這項分析中會看到該門檻對癡呆症發病率的影響。

Third, we conducted the identical comparison of individuals around the 2 September 1933 date-of-birth threshold as in our primary analysis, except for starting the follow-up period 7 years before the start of the zoster vaccine program rollout. If there was an intervention that used the 2 September 1933 date-of-birth eligibility threshold but was implemented before the rollout of the zoster vaccine program, then we may expect to see an effect of the September 1933 threshold on dementia incidence in this analysis.

第四，我們驗證了我們在分析中觀察到的癡呆發病率的影響似乎是特定於癡呆的。如果一個使用與帶狀皰疹疫苗計劃完全相同的出生日期資格門檻的干預確實存在，那麼它不太可能僅影響癡呆風險，而不會對其他健康結果產生影響。因此，我們進行了與使用癡呆發病率作為結果時相同的分析，但針對 2019 年威爾士 70 歲以上年齡組的十個主要殘疾調整生命年和死亡原因，以及所有屬於查爾森共病指數的條件。

Fourth, we verified that the effects that we observed in our analyses for dementia incidence appear to be specific to dementia. If an intervention that used the exact same date-of-birth eligibility threshold as the zoster vaccine program indeed existed, it would be unlikely to only affect dementia risk without also having an influence on other health outcomes. We therefore conducted the same analysis as for when using dementia incidence as the outcome but for each of the ten leading causes of disability-adjusted life years and mortality in Wales in 2019 for the age group 70+ years³³, as well as all conditions that are part of the Charlson Comorbidity Index³⁴.

第五，我們使用 2011 年威爾斯人口普查的數據，檢測在 1933 年 9 月 2 日出生日期的門檻下教育成就的斷裂。如果某項教育政策使用了 1933 年 9 月 2 日（或特定的 1933 年 9 月 2 日）出生日期的門檻，並且該政策在提高教育成就方面有效，那麼我們就會期望在 1933 年 9 月 2 日的門檻下，合格和不合格個體之間的教育水平存在斷裂。我們對這個平衡測試使用了與 SAIL 數據庫中的主要分析相同的分析方法，只是我們根據 Armstrong 和 Kolesár 的方法計算了“誠實”的置信區間，因為這些數據中的分配變量（出生月份）是按月的，因此比我們在 SAIL 數據庫中的分析中的分配變量（出生周）更粗糙。這種方法可以防止潛在的模型錯誤規範和由此導致的使用更標準方法計算的置信區間的低覆蓋率。 這些誠實的信心區間是保守的，因為它們具有良好的覆蓋性質，無論回歸不連續性分析中的函數形式是否錯誤指定，只要真實的函數形式落在某個函數類別內。對於這個類別，我們考慮了一個由條件期望函數的二階導數界限定義的函數類別，該函數將出生日期映射到達到某一教育水平的概率。我們通過依賴於 Armstrong 和Kolesár ⁸⁸ 提出的高階多項式的全局估計，使用了各自曲率的保守界限。

Fifth, we tested for discontinuities in educational attainment at the 2 September 1933 date-of-birth threshold using data from the 2011 census in Wales⁷⁴. If an educational policy had used a 2 September (or specifically 2 September 1933) date-of-birth threshold and the policy was effective in increasing educational attainment, we would then expect discontinuities at the 2 September 1933 threshold in the attained education level between eligible and ineligible individuals. We used the identical analysis approach for this balance test as for our primary analyses in the SAIL database, except that we computed ‘honest’ confidence intervals based on the approach by Armstrong and Kolesár because the assignment variable (month of birth) in these data was monthly, and therefore coarser than the assignment variable (week of birth) in our analyses in the SAIL database^87,88. This approach guards against potential vulnerability to model misspecification and resulting under-coverage of confidence intervals computed with more standard methods. These honest confidence intervals are conservative in the sense that they have good coverage properties irrespective of whether the functional form in the regression discontinuity analysis is misspecified, provided that the true functional form falls within a certain class of functions. For this class, we considered a function class defined by bounds on the second derivative of the conditional expectation function mapping date of birth to the probability attaining a certain educational level. We used conservative bounds of the respective curvatures by relying on global estimation of higher-order polynomials as proposed by Armstrong and Kolesár⁸⁸.

對不同分析方法的穩健性
Robustness to a different analytical approach

我們另外使用了差異中的差異工具變量方法（DID-IV）來確認我們的回歸不連續設計的發現，因為與回歸不連續分析相比，這種方法不依賴於連續性假設（即，潛在混雜變量不會在帶狀皰疹疫苗計劃的出生日期資格閾值上突然改變的假設）。為此，我們將樣本限制在 1926 年 3 月 1 日至 1934 年 2 月 28 日之間出生的患者。這個樣本由 96,767 名成年人組成，其中 7,752 名（8.0%）符合帶狀皰疹疫苗接種的資格，3,949 名（4.1%）實際接種了帶狀皰疹疫苗。然後，我們將樣本分為圍繞 9 月 1 日的年度隊列（即，一個隊列是所有在某一年 3 月 1 日至下一年 2 月 28 日之間出生的患者）。最後，我們將每個年度隊列分為 9 月前出生季節和 9 月後出生季節。使用差異中的差異方法，我們然後比較了在 9 月前和 9 月後出生季節的患者之間以及年度隊列之間的結果（新診斷的癡呆症）。 更精確地說，我們測試了 1933/1934 年出生的隊列與其他隊列之間，出生季節對結果的影響差異是否不同。在這樣做的過程中，我們利用了這樣一個事實：在 1933/1934 年出生的隊列中，帶狀皰疹疫苗接種的資格僅在兩個出生季節之間存在差異，而在其他隊列中則不存在，同時考慮到 9 月前和 9 月後的出生季節可能因其他原因而系統性地不同。

We additionally used a difference-in-differences instrumental variable approach (DID-IV) to confirm the findings from our regression discontinuity design because, in contrast to the regression discontinuity analysis, this approach does not rely on the continuity assumption (that is, the assumption that potential confounding variables do not abruptly change at precisely the date-of-birth eligibility threshold for the zoster vaccine program). To do so, we restricted our sample to patients born between 1 March 1926 and 28 February 1934. This sample consists of 96,767 adults, of whom 7,752 (8.0%) were eligible for, and 3,949 (4.1%) actually received, zoster vaccination. We then divided our sample into yearly cohorts centred around 1 September (that is, a cohort is all patients born between 1 March of one year and 28 February of the following year). Finally, we divided each yearly cohort into a pre-September birth season and a post-September birth season. Using a difference-in-differences approach, we then compared the outcome (new diagnoses of dementia) between patients born in pre- and post-September birth seasons and across yearly cohorts. More precisely, we tested whether the difference in outcomes across birth seasons is different for the 1933/1934 cohort than for the other cohorts. In doing so, we exploit the fact that zoster vaccination eligibility only differs between the two birth seasons in the 1933/1934-cohort but not in other cohorts, while accounting for the possibility that pre-September and post-September birth seasons may be systematically different for other reasons.

我們的差異中的差異設置意味著，九月後出生季節指標與 1933/1934 年出生隊列指標之間的互動構成了接種帶狀皰疹疫苗的工具變量，使我們能夠估計 CACE（即，實際接種疫苗的合規者的效果）。這種 DID-IV 方法依賴於兩個重要假設。根據 IV 分析的標準排除限制假設，我們的 DID-IV 方法的 IV 組件假設疫苗資格僅通過實際接種疫苗的變化影響結果。我們的 DID-IV 方法的 DID 組件假設，在沒有疫苗資格規則的情況下，1933/1934 年隊列與其他隊列之間的疫苗接種率和癡呆發病率的季節差異將是相同的。為了檢查這一假設的有效性，我們繪製了按出生隊列和出生季節的平均疫苗接種率和癡呆發病率及 95%置信區間（補充圖 29）。正如預期的那樣，我們發現疫苗接種率的出生季節之間的差異僅在 1933/1934 年出生隊列中出現。 其他出生隊列中缺乏出生季節之間的差異支持了我們的 DID 假設的有效性。

Our difference-in-differences setup implies that the interaction between the post-September birth season indicator and the 1933/1934-cohort indicator constitutes an instrumental variable for receipt of the zoster vaccine, enabling us to estimate the CACE (that is, the effect of actually receiving the vaccine among the compliers). This DID-IV approach relies on two important assumptions. As per the standard exclusion restriction assumption of IV analyses, the IV component of our DID-IV approach assumes that vaccine eligibility affects the outcome solely through a change in actual vaccine receipt. The DID component of our DID-IV approach assumes that, in the absence of the vaccine eligibility rule, the between-birth-season difference in vaccine uptake and in dementia incidence would have been the same in the 1933/1934 cohort as in the other cohorts. To investigate the validity of this assumption, we plotted the mean vaccine uptake and dementia incidence with 95% CIs by birth cohorts and birth seasons (Supplementary Fig. 29). As expected, we find that the between-birth-season differences in vaccine uptake diverge only in the 1933/1934 birth cohort. The absence of a between-birth-season difference in other birth cohorts supports the validity of our DID assumption.

在這個 DID-IV 框架中，為了估計 CACE，我們使用了兩階段最小二乘回歸。在第一階段，我們通過以下回歸方程識別由於外生變化的疫苗接種資格所導致的疫苗接種率：

To estimate the CACE in this DID-IV framework, we used two-stage least-squares regression. In the first stage, we identify the vaccine uptake due to the exogeneous change in vaccination eligibility by the following regression equation:

其中 V 是一個二元變數，表示患者 i 實際上接種了帶狀皰疹疫苗。S 和 C 是二元變數，分別表示患者 i 出生於九月後的出生季節和 1933/1934 出生隊列。γ 確定由於資格變更而導致的疫苗接種率。θ、η _m 和 η _c 分別是常數項、出生月份（1 月、2 月、……、12 月）和出生隊列（1926/1927、1927/1928、……、1933/1934）的固定效應。ϵ 是誤差項。
where V_i is a binary variable indicating patient i actually received the zoster vaccine. S_i and C_i are binary variables indicating that patient i is born in the post-September birth season and in the 1933/1934 birth cohort, respectively. γ identifies the vaccine uptake due to the change in eligibility. θ, η_m and η_c are the constant term, birth month (January, February, …, December) and birth cohort (1926/1927, 1927/1928, …, 1933/1934) fixed effect, respectively. ϵ_i is the error term.

在第二階段，我們通過以下回歸估計疫苗接種的效果：
In the second stage, we estimate the effect of vaccine receipt by the following regression:

其中 Y 是患者 i 的結果。 V^i 是從第一階段回歸 (4) 預測的疫苗接種概率。係數 β 確定了 CACE。α、η _m 和 η _c 分別是常數項、出生月份和出生隊列 (1926/1927、1927/1928、……、1933/1934) 的固定效應。ϵ 是誤差項。
where Y_i is the outcome of patient i. V^i is the probability of vaccine receipt predicted from the first-stage regression (4). The coefficient β identifies the CACE. α, η_m and η_c are the constant term, birth month and birth cohort (1926/1927, 1927/1928, …, 1933/1934) fixed effect, respectively. ϵ_i is the error term.

對不同分析規範的穩健性檢查
Robustness checks to different analytical specifications

我們進行了一系列額外的穩健性檢查。首先，我們不是在 2013 年 9 月 1 日為所有個體開始隨訪期，而是調整了隨訪期，以考慮計劃的分階段推出，通過在每個個體首次符合帶狀皰疹疫苗接種資格的日期開始隨訪期（如“威爾士帶狀皰疹疫苗推出的描述”部分所述）（補充圖 30）。在這些分析中，我們控制了隊列固定效應，以考慮在這個移動隨訪窗口開始的日曆年中，根據計劃年份，隊列之間的差異為一到兩年。我們為不符合資格的個體和第一批補救隊列定義了一個隊列固定效應，然後為每組同時符合資格的患者包括額外的隊列固定效應。其次，我們通過將癡呆症定義為新的多奈哌齊、加蘭他敏、利伐斯的明或美金剛鹽酸鹽的處方來變更我們對新癡呆症診斷的定義。 第三，我們測試了符合帶狀皰疹疫苗接種資格對新診斷癡呆症、帶狀皰疹和後帶狀皰疹神經痛的影響結果是否在 0、2、4、6、8、10 和 12 個月的寬限期內保持一致（即，自指標日期起的時間段，在此之後的隨訪時間被認為開始，以便考慮在疫苗接種後發展出完整免疫反應所需的時間）（補充圖 31）。第四，我們展示了使用 0.25、0.50、0.75、1.00、1.25、1.50、1.75 和 2.00 倍的 MSE 最佳帶寬的帶寬選擇下的結果（補充圖 32）。第五，我們驗證了使用局部二次多項式規範而不是局部線性回歸時，我們的結果是相似的。

We conducted a series of additional robustness checks. First, instead of starting the follow-up period for all individuals on 1 September 2013, we adjusted the follow-up period to account for the staggered rollout of the program by beginning the follow-up period for each individual on the date on which they first became eligible for the zoster vaccine (as described in the ‘Description of the zoster vaccine rollout in Wales’ section) (Supplementary Fig. 30). We controlled for cohort fixed effects in these analyses to account for the one- to two-year (depending on the year of the program) differences between cohorts in the calendar year in which this moving follow-up window started. That is, we defined one cohort fixed effect for ineligible individuals and the first catch-up cohort and then included additional cohort fixed effects for each group of patients who became eligible at the same time. Second, we varied our definition of a new diagnosis of dementia by implementing our analysis when defining dementia as a new prescription of donepezil hydrochloride, galantamine, rivastigmine or memantine hydrochloride. Third, we tested whether our results for the effect of being eligible for zoster vaccination on new diagnoses of dementia, shingles and postherpetic neuralgia hold across grace periods (that is, time periods since the index date after which follow-up time is considered to begin to allow for the time needed for a full immune response to develop after vaccine administration) of 0, 2, 4, 6, 8, 10 and 12 months (Supplementary Fig. 31). Fourth, we show our results with bandwidth choices of 0.25, 0.50, 0.75, 1.00, 1.25, 1.50, 1.75 and 2.00 times the MSE-optimal bandwidth (Supplementary Fig. 32). Fifth, we verified that our results are similar when using a local second-order polynomial specification instead of local linear regression.

分析以探討效應機制
Analyses to explore the effect mechanism

我們進行了四項分析以檢查這一潛在的影響機制，其中前三項在主文中詳細描述。第四項分析是一項事件研究，專注於隨訪期間帶狀皰疹診斷的日期。我們的事件研究比較了每個月份的平均結果與帶狀皰疹診斷日期前一個月的結果。我們的回歸模型使用月份固定效應控制時間變化（例如，由於研究人群的老化或醫療提供者就診的季節性模式）。

We conducted four analyses to examine this potential effect mechanism, the first three of which are described in detail in the main text. The fourth analysis was an event study that focused on the date of a shingles diagnosis during the follow-up period. Our event study compared the mean outcome in each month relative to the month before the date of the shingles diagnosis. Our regression model controls for changes over time (such as due to ageing of the study population or seasonal patterns in healthcare provider visits) using month-level fixed effects.

為了實施我們的事件研究，我們將研究人群限制在 56,098 名出生於我們主要回歸不連續性分析的最佳帶寬內的個體。我們隨後將事件級數據匯總為每月的縱向數據，涵蓋 2013 年 9 月至 2022 年 3 月。對於每個感興趣的結果（如主文中所述），我們隨後估計了以下事件研究回歸：

To implement our event study, we restricted our study population to those 56,098 individuals born within the MSE-optimal bandwidth of our primary regression discontinuity analysis for dementia. We then aggregated our event-level data into monthly longitudinal data, spanning September 2013 to March 2022. For each outcome of interest (as described in the main text), we then estimated the following event-study regression:

其中 Y _it 是個體 i 在時期 t 的關注結果；帶狀皰疹是一個二元變數，如果個體在隨訪期間被診斷為帶狀皰疹則等於一；D _k 是在帶狀皰疹診斷前後 k 個月的指示變數（k = −36, −35,…, 35, 36，對於在隨訪期間從未被診斷為帶狀皰疹的個體設為零）；γ _k 是關注的係數，捕捉在帶狀皰疹診斷前一個月相對於第 k 個月的結果差異；η 是捕捉個體之間時間不變差異的個體層級固定效應；λ _t 是捕捉不同時期差異的月份層級固定效應。我們使用了允許在個體層級聚類的標準誤，因此也考慮了自相關。

where Y_it is the outcome of interest for individual i in period t; shingles is a binary variable equal to one if the individual was diagnosed with shingles during the follow-up period; D_k are indicator variables for the k months before and after the shingles diagnosis (with k = −36, −35,…, 35, 36, and set to zero for individuals who were never diagnosed with shingles during the follow-up period); γ_k are the coefficients of interest, which capture the difference in the outcome in month k relative to the month before the shingles diagnosis; η_i is an individual-level fixed effect capturing time-invariant differences across individuals; and λ_t is a month-level fixed effect, capturing differences across periods. We used standard errors that allowed for clustering at the individual level, and therefore for autocorrelation.

我們進行了四項分析以檢查這一潛在的效應機制。首先，我們實施了與我們主要分析相同的回歸不連續性，只是我們將在隨訪期間被診斷為帶狀皰疹的二元變量作為協變量納入。為了使結果估計成為帶狀皰疹疫苗對癡呆症發病率影響的無偏測量，必須沒有與新的癡呆症診斷和被診斷為帶狀皰疹的概率相關的變量（即，沒有中介與結果之間關係的混淆） ⁸⁹ 。

We conducted four analyses to examine this potential effect mechanism. First, we implemented the identical regression discontinuity as in our primary analysis, except that we included a binary variable for being diagnosed with shingles during the follow-up period as a covariate. For the resulting estimate to be an unbiased measure of the degree to which the effect of zoster vaccination on dementia incidence is mediated by shingles diagnoses, there must be no variables that are related to both new dementia diagnoses and the probability of being diagnosed with shingles (that is, no confounding of the mediator-to-outcome relationship)⁸⁹.

其次，為了檢查帶狀皰疹疫苗接種對癡呆症延遲或預防效果在隨訪期間何時開始顯現，我們繪製了卡普蘭-邁爾圖和（為了考慮競爭風險）累積發生率曲線，對象為接近 1933 年 9 月 2 日出生的個體。這項分析基於局部隨機化的概念 ^28,29 ，該概念依賴於在 1933 年 9 月 2 日之前和之後立即出生的個體之間的可交換性。為了定義我們分析的帶寬，在這個帶寬內我們可以合理地假設在閾值上存在可交換性，同時最大化統計功效，我們使用了在基線人口統計和臨床特徵上達到平衡的最寬帶寬，該特徵適用於符合和不符合帶狀皰疹疫苗接種資格的個體。我們在主要回歸不連續性分析中評估了從 100%到 10%的 MSE 最佳帶寬（90.6 週）的帶寬，減少幅度為 10%。我們用於平衡測試的變量是補充圖中列出的 14 個變量。 1 和 2（除了過去乳腺癌篩檢、乳腺癌和前列腺癌診斷的性別特定變數）使用顯著性閾值 P < 0.05，同時使用 Benjamini–Hochberg 程序控制假陽性率 ⁹⁰ 。達到所有變數平衡的最大帶寬為 54.4 週。

Second, to examine when during the follow-up period the dementia-delaying or dementia-preventing effect of zoster vaccination begins to emerge, we plotted Kaplan–Meier plots and (to account for competing risks) cumulative incidence curves among individuals born close to 2 September 1933. This analysis was based on the concept of local randomization^28,29, which relies on exchangeability of individuals born immediately before versus immediately after 2 September 1933. To define the bandwidth for our analysis in which we could reasonably assume exchangeability across the threshold while maximizing statistical power, we used the widest bandwidth for which we achieved balance in baseline demographic and clinical characteristics of individuals eligible versus ineligible for zoster vaccination. We evaluated bandwidths ranging from 100% to 10% of the MSE-optimal bandwidth (90.6 weeks) in our primary regression discontinuity analysis in 10% decrements. The variables we used for our balance tests were the 14 variables listed in Supplementary Figs. 1 and 2 (except for the more sex-specific variables of past breast cancer screening, breast cancer and prostate cancer diagnoses) using a significance threshold of P < 0.05, while controlling for the false-discovery rate using the Benjamini–Hochberg procedure⁹⁰. The largest bandwidth that achieved balance across all variables was 54.4 weeks.

第三，我們調查在帶狀皰疹發作期間接受抗病毒治療是否與相對於未接受治療的痴呆風險降低相關，我們將研究對象限制為在 2000 年 1 月 1 日之後的任何時間被診斷為帶狀皰疹且在 2000 年 1 月 1 日之前未被診斷為痴呆的個體。我們在此分析中的關注點是個體是否在首次帶狀皰疹診斷後的三個月內獲得了抗病毒藥物（阿昔洛韋、法莫克洛韋、伐昔洛韋或腺苷普那韋）的處方。個體的隨訪時間從首次帶狀皰疹診斷之日起，直到死亡日期、搬出威爾士、家庭醫生註銷或數據可用性結束（2022 年 3 月 1 日）。然後，我們使用多變量 Cox 比例風險模型將首次記錄的帶狀皰疹發作日期之後的痴呆診斷回歸到患者是否為首次帶狀皰疹發作獲得抗病毒藥物處方。 在穩健性檢查中，我們要求新的癡呆症診斷必須在第一次帶狀皰疹診斷日期後至少 12 個月才可確立。我們對回歸進行了性別調整，並對第一次帶狀皰疹感染的年齡進行了限制性立方樣條（有三個結），以及補充圖 1 中的 12 個變數（不包括過去的乳腺癌和前列腺癌診斷）。

Third, to investigate whether antiviral treatment during a shingles episode was associated with a reduction in the risk of dementia relative to not receiving treatment during a shingles episode, we restricted our study population to those individuals who received a diagnosis of shingles at any time after 1 January 2000 and had not received a diagnosis of dementia before 1 January 2000. Our exposure of interest in this analysis was whether or not an individual received a prescription of antiviral medication (acyclovir, famcyclovir, valacyclovir or inosine pranobex) within three months of the first shingles diagnosis. Individuals were followed up from the date of first shingles diagnosis until either the date of death, moving out of Wales, GP deregistration or end of data availability (1 March 2022). We then used a multivariable Cox proportional hazards model to regress diagnoses of dementia made after the date of the first recorded shingles episode onto whether or not the patient received an antiviral medication prescription for the first shingles episode. In a robustness check, we required that a new diagnosis of dementia must have been made at least 12 months after the date of the first shingles diagnosis. We adjusted our regressions for gender, restricted cubic splines (with three knots) of age at the first shingles infection, and the 12 variables in Supplementary Fig. 1 (excluding past breast and prostate cancer diagnoses).

第四，我們探討經歷反覆帶狀皰疹發作是否與比僅有單次發作更高的癡呆風險相關，使用的研究人群與我們對治療與未治療帶狀皰疹的分析相同。我們通過 1:1 傾向得分匹配，將在 2000 年 1 月 1 日後有多於一次帶狀皰疹診斷（診斷日期必須相隔至少三個月）的個體與在 2000 年 1 月 1 日後僅接受一次帶狀皰疹診斷的個體進行匹配。我們在首次帶狀皰疹診斷的日期接近度以及與我們對治療與未治療帶狀皰疹分析相同的基線變數列表上進行匹配，並強制在出生週和性別上進行精確匹配。在每對匹配中，我們使用多次帶狀皰疹診斷的個體的第二次帶狀皰疹診斷日期作為隨訪期間的開始日期。然後，我們使用 Cox 比例風險模型，將隨訪期間新診斷的癡呆與個體是否接受過多次帶狀皰疹診斷進行回歸分析。 在穩健性檢查中，我們再次要求新的癡呆症診斷必須在隨訪期開始日期後至少 12 個月內做出。

Fourth, to explore whether experiencing recurrent shingles episodes was associated with a higher risk of dementia than having only a single episode, we used the same study population as in our analysis for treated versus untreated shingles. We matched individuals (via 1:1 propensity score matching) who had more than one shingles diagnosis (with the diagnosis dates having to be at least three months apart) after 1 January 2000 to individuals who only received a single shingles diagnosis after 1 January 2000. We matched individuals on proximity in the date of their first shingles diagnosis as well as the same list of baseline variables as for our analysis of treated versus untreated shingles, and forced an exact match on week of birth and gender. In each matched pair, we used the date of the second shingles diagnosis of the individual with more than one shingles diagnosis as the start date of the follow-up period. Using a Cox proportional hazards model, we then regressed new diagnoses of dementia made during the follow-up period onto whether or not the individual had received more than one shingles diagnosis. In a robustness check, we again required that a new diagnosis of dementia must have been made at least 12 months after the start date of the follow-up period.

為了估計本節中主文所描述的治療效果異質性，我們將模糊回歸不連續模型與一個二元變數完全交互，該變數指示是否具有相關條件（例如，自身免疫性疾病）。具體而言，完全交互模型的規範為：
To estimate the treatment effect heterogeneities described under this section in the main text, we fully interacted our fuzzy regression discontinuity model with a binary variable that indicates having the condition in question (for example, an autoimmune condition). Precisely, the fully interacted model was specified as:

其中下標 i 索引個體。Y 是一個二元變量，如果個體在隨訪期間被新診斷為癡呆，則等於 1。二元變量 V 表示接種帶狀皰疹疫苗。二元變量 D 表示符合接種帶狀皰疹疫苗的資格（即，出生於 1933 年 9 月 2 日或之後）。術語 WOB − c ₀ 表示個體的出生週，圍繞著出生資格門檻的日期。交互項 D × (WOB − c ₀ ) 允許回歸線的斜率在出生資格門檻的兩側有所不同。二元變量 HET 等於 1，如果個體有相關條件。添加項 (WOB − c ₀ ) × HET 和 D × (WOB − c ₀ ) × HET 允許斜率根據此條件變化。

where the subscript i indexes individuals. Y_i is a binary variable equal to 1 if an individual was newly diagnosed with dementia during the follow-up period. The binary variable V_i indicates receipt of the zoster vaccine. The binary variable D_i indicates eligibility for the zoster vaccine (that is, born on or after 2 September 1933). The term WOB_i − c₀ indicates an individual’s week of birth centred around the date-of-birth eligibility threshold. The interaction term D_i × (WOB_i − c₀) allows for the slope of the regression line to differ on either side of the date-of-birth eligibility threshold. The binary variable HET_i is equal to one if an individual had the condition in question. Adding the terms (WOB_i − c₀) × HET_i and D_i × (WOB_i − c₀) × HET_i allows the slopes to vary by this condition.

V 和 V × HET 由 D 和 D × HET 作為工具變量。使用兩階段最小二乘法，參數 β ₄ 確定了效應異質性，即有病和無病患者在結果上的 CACE 差異。β ₁ 和 β ₁ + β ₄ 分別確定了參考組和比較組中遵從者的效應。效應和異質性的估計以絕對值報告。為了與我們的主要模糊回歸不連續模型（即不包含 HET 和交互項）保持一致，我們使用了局部線性三角核回歸和來自各自結果的主要模型的 MSE 最優帶寬。

V_i and V_i × HET_i are instrumented by D_i and D_i × HET_i. Using the two-stage least-squares approach, the parameter β₄ identifies the effect heterogeneity, that is, the difference in CACE on the outcome between patients with and without the condition. β₁ and β₁ + β₄ identify the effect among compliers in the reference and comparison group, respectively. The estimates of the effects and heterogeneity are reported in absolute terms. To be consistent with our primary fuzzy regression discontinuity model (that is, without the HET_i and interaction terms), we used local linear triangular kernel regressions and the MSE-optimal bandwidth from the primary model of the respective outcome.

在我們對自體免疫和過敏狀況的分析中，我們使用了之前定義的 19 種最常見的自體免疫狀況，並將其中 11 種較不常見的狀況歸類為罕見狀況類別。我們判斷那些在我們的隊列中隨訪期間發生率低於 1%的狀況為罕見狀況。這些罕見狀況包括阿迪森氏病、僵直性脊椎炎、乳糜瀉、橋本氏甲狀腺炎、多發性硬化症、重症肌無力、原發性膽汁性肝硬化、乾燥症、系統性紅斑狼瘡、系統性硬化症和白癜風。對於常見的過敏狀況，我們使用了之前定義的那些。

For our analyses for autoimmune and allergic conditions, we used the 19 most common autoimmune conditions as defined previously⁹¹, and grouped the 11 least common conditions among them into a rare conditions category. We judged those conditions to be rare that had an incidence of less than 1% during the follow-up period in our cohort. These rare conditions included Addison’s disease, ankylosing spondylitis, Coeliac disease, Hashimoto’s thyroiditis, multiple sclerosis, myasthenia gravis, primary biliary cirrhosis, Sjögren’s syndrome, systemic lupus erythematosus, systemic sclerosis and vitiligo. For common allergic conditions, we used those defined previously⁹².

參考文獻 References

1. Wainberg, M. et al. 病毒假說：疱疹病毒如何可能促進阿茲海默症。分子精神醫學 26, 5476–5480 (2021)。
   Wainberg, M. et al. The viral hypothesis: how herpesviruses may contribute to Alzheimer’s disease. Mol. Psychiatry 26, 5476–5480 (2021).

   Article PubMed PubMed Central Google Scholar 

2. Devanand, D. P. 病毒假說與阿茲海默症的抗病毒治療。當前神經學與神經科學報告 18, 55 (2018)。
   Devanand, D. P. Viral hypothesis and antiviral treatment in Alzheimer’s disease. Curr. Neurol. Neurosci. Rep. 18, 55 (2018).

   Article CAS PubMed PubMed Central Google Scholar 

3. Moir, R. D., Lathe, R. & Tanzi, R. E. 阿茲海默病的抗微生物保護假說。阿茲海默症與癡呆症。14, 1602–1614 (2018)。
   Moir, R. D., Lathe, R. & Tanzi, R. E. The antimicrobial protection hypothesis of Alzheimer’s disease. Alzheimers Dement. 14, 1602–1614 (2018).

   Article PubMed Google Scholar 

4. Itzhaki, R. F. 等。微生物與阿茲海默症。J. Alzheimers Dis. 51, 979–984 (2016)。
   Itzhaki, R. F. et al. Microbes and Alzheimer’s disease. J. Alzheimers Dis. 51, 979–984 (2016).

   Article PubMed PubMed Central Google Scholar 

5. Hyde, V. R. 等。抗疱疹 tau 通過 cGAS-STING-TBK1 途徑保護阿茲海默症中的神經元。Cell Rep. 44, 115109 (2025)。
   Hyde, V. R. et al. Anti-herpetic tau preserves neurons via the cGAS-STING-TBK1 pathway in Alzheimer’s disease. Cell Rep. 44, 115109 (2025).

6. Benn, C. S., Fisker, A. B., Rieckmann, A., Sørup, S. & Aaby, P. 疫苗學：是時候改變範式了嗎？柳葉刀傳染病 20, e274–e283 (2020)。
   Benn, C. S., Fisker, A. B., Rieckmann, A., Sørup, S. & Aaby, P. Vaccinology: time to change the paradigm? Lancet Infect. Dis. 20, e274–e283 (2020).

   Article CAS PubMed Google Scholar 

7. Benn, C. S. 等人。非特異性效應對疫苗測試、批准和監管的影響。藥物安全性。46，439–448（2023 年）。
   Benn, C. S. et al. Implications of non-specific effects for testing, approving, and regulating vaccines. Drug Saf. 46, 439–448 (2023).

   Article PubMed PubMed Central Google Scholar 

8. Aaby, P. 等人。疫苗的非特異性和性別差異效應。自然免疫學評論。20，464–470（2020 年）。
   Aaby, P. et al. The non-specific and sex-differential effects of vaccines. Nat. Rev. Immunol. 20, 464–470 (2020).

   Article CAS PubMed PubMed Central Google Scholar 

9. McGovern, M. E. & Canning, D. 1985 年至 2011 年疫苗接種與全因兒童死亡率：來自人口與健康調查的全球證據。美國流行病學雜誌。182，791–798（2015 年）。
   McGovern, M. E. & Canning, D. Vaccination and all-cause child mortality from 1985 to 2011: global evidence from the Demographic and Health Surveys. Am. J. Epidemiol. 182, 791–798 (2015).

   Article PubMed PubMed Central Google Scholar 

10. Michalik, F. 等人。帶狀皰疹疫苗對英格蘭和威爾士因癡呆症死亡發生率的影響。預印本於 medRxiv https://doi.org/10.1101/2023.09.08.23295225 (2023)。
    Michalik, F. et al. The effect of herpes zoster vaccination on the occurrence of deaths due to dementia in England and Wales. Preprint at medRxiv https://doi.org/10.1101/2023.09.08.23295225 (2023).

11. Wu, X. et al. 成人疫苗接種作為預防癡呆的保護因素：基於人群的觀察性研究的元分析和系統評價。Front. Immunol. 13, 872542 (2022).
    Wu, X. et al. Adult vaccination as a protective factor for dementia: a meta-analysis and systematic review of population-based observational studies. Front. Immunol. 13, 872542 (2022).

    Article CAS PubMed PubMed Central Google Scholar 

12. Hernan, M. A. & Robins, J. M. 因果推斷：如果 (Chapman & Hall/CRC, 2020)。
    Hernan, M. A. & Robins, J. M. Causal Inference: What If (Chapman & Hall/CRC, 2020).

13. Simonsen, L., Taylor, R. J., Viboud, C., Miller, M. A. & Jackson, L. A. 流感疫苗對老年人死亡率的益處：一個持續的爭議。Lancet Infect. Dis. 7, 658–666 (2007)。
    Simonsen, L., Taylor, R. J., Viboud, C., Miller, M. A. & Jackson, L. A. Mortality benefits of influenza vaccination in elderly people: an ongoing controversy. Lancet Infect. Dis. 7, 658–666 (2007).

    Article PubMed Google Scholar 

14. 帶狀皰疹：指導和疫苗接種計劃 www.gov.uk/government/collections/shingles-vaccination-programme (英國健康安全局，2021)。
    Shingles: Guidance and Vaccination Programme www.gov.uk/government/collections/shingles-vaccination-programme (UK Health Security Agency, 2021).

15. Harris, K. 等。常規疫苗接種對 65 歲及以上人群阿茲海默症風險的影響：一項基於索賠的隊列研究，使用傾向得分匹配。J. Alzheimers Dis. 95, 703–718 (2023)。
    Harris, K. et al. The impact of routine vaccinations on Alzheimer’s disease risk in persons 65 years and older: a claims-based cohort study using propensity score matching. J. Alzheimers Dis. 95, 703–718 (2023).

    Article CAS PubMed PubMed Central Google Scholar 

16. Lophatananon, A. 等。帶狀皰疹、Zostavax 疫苗接種與發展癡呆風險：來自英國生物銀行隊列的嵌套病例對照研究結果。BMJ Open 11, e045871 (2021)。
    Lophatananon, A. et al. Shingles, Zostavax vaccination and risk of developing dementia: a nested case-control study-results from the UK Biobank cohort. BMJ Open 11, e045871 (2021).

    Article PubMed PubMed Central Google Scholar 

17. Scherrer, J. F. 等人。帶狀皰疹疫苗對新發癡呆症的影響：對兩個患者隊列的回顧性研究。PLoS ONE 16, e0257405 (2021)。
    Scherrer, J. F. et al. Impact of herpes zoster vaccination on incident dementia: A retrospective study in two patient cohorts. PLoS ONE 16, e0257405 (2021).

    Article CAS PubMed PubMed Central Google Scholar 

18. Schnier, C.、Janbek, J.、Lathe, R. 和 Haas, J. 在威爾士 2013 年至 2020 年間，水痘帶狀皰疹疫苗接種後癡呆症發病率降低。阿爾茨海默病與癡呆症 8, e12293 (2022)。
    Schnier, C., Janbek, J., Lathe, R. & Haas, J. Reduced dementia incidence after varicella zoster vaccination in Wales 2013–2020. Alzheimers Dement. 8, e12293 (2022).

    Google Scholar 

19. Wiemken, T. L. 等人。比較接受兩次、一次或未接種疫苗的老年人癡呆症發病率。美國老年醫學會雜誌 70, 1157–1168 (2022)。
    Wiemken, T. L. et al. Comparison of rates of dementia among older adult recipients of two, one, or no vaccinations. J. Am. Geriatr. Soc. 70, 1157–1168 (2022).

    Article PubMed Google Scholar 

20. Lophatananon, A. 等人。帶狀皰疹和流感疫苗接種與發展癡呆風險的關聯：基於英國臨床實踐研究數據庫的人口基礎隊列研究。BMC 公共衛生 23, 1903 (2023)。
    Lophatananon, A. et al. The association of herpes zoster and influenza vaccinations with the risk of developing dementia: a population-based cohort study within the UK Clinical Practice Research Datalink. BMC Publ. Health 23, 1903 (2023).

    Article Google Scholar 

21. Ukraintseva, S. 等人。感染和疫苗與阿茲海默症的關聯指向免疫力受損而非特定病原體在阿茲海默症中的作用。實驗老年學 190, 112411 (2024)。
    Ukraintseva, S. et al. Associations of infections and vaccines with Alzheimer’s disease point to a role of compromised immunity rather than specific pathogen in AD. Exp. Gerontol. 190, 112411 (2024).

    Article CAS PubMed PubMed Central Google Scholar 

22. Lehrer, S. & Rheinstein, P. H. 帶狀皰疹疫苗可降低癡呆風險。In Vivo 35, 3271–3275 (2021)。
    Lehrer, S. & Rheinstein, P. H. Herpes zoster vaccination reduces risk of dementia. In Vivo 35, 3271–3275 (2021).

    Article CAS PubMed PubMed Central Google Scholar 

23. Wilkinson, T. 等人。藥物處方與癡呆發病率：針對 17,000 例癡呆病例的藥物廣泛關聯研究，參與者達 50 萬。流行病學與社區健康雜誌 76, 223–229 (2022)。
    Wilkinson, T. et al. Drug prescriptions and dementia incidence: a medication-wide association study of 17000 dementia cases among half a million participants. J. Epidemiol. Community Health 76, 223–229 (2022).

    Article PubMed Google Scholar 

24. Tang, E., Ray, I., Arnold, B. F. & Acharya, N. R. 重組帶狀皰疹疫苗與癡呆風險。疫苗 46, 126673 (2024)。
    Tang, E., Ray, I., Arnold, B. F. & Acharya, N. R. Recombinant zoster vaccine and the risk of dementia. Vaccine 46, 126673 (2024).

    Article PubMed PubMed Central Google Scholar 

25. Shingrix 疫苗的引入及擴大符合條件的群體計劃信函 https://www.gov.uk/government/publications/shingles-vaccination-programme-changes-from-september-2023-letter/introduction-of-shingrix-vaccine-for-the-whole-programme-and-expansion-of-eligible-cohorts-letter (英國健康安全局, 2023)。
    Introduction of Shingrix Vaccine for the Whole Programme and Expansion of Eligible Cohorts Letter https://www.gov.uk/government/publications/shingles-vaccination-programme-changes-from-september-2023-letter/introduction-of-shingrix-vaccine-for-the-whole-programme-and-expansion-of-eligible-cohorts-letter (UK Health Security Agency, 2023).

26. SAIL 數據庫 saildatabank.com/（斯旺西大學，2025 年）。
    SAIL Databank saildatabank.com/ (Swansea University, 2025).

27. 歸因數據集 GP 註冊人口按 ONS 人口估計縮放—2011 digital.nhs.uk/data-and-information/publications/statistical/attribution-dataset-gp-registered-populations/attribution-data-set-gp-registered-populations-scaled-to-ons-population-estimates-2011（NHS 數字，2012 年）。
    Attribution Data Set GP-Registered Populations Scaled to ONS Population Estimates—2011 digital.nhs.uk/data-and-information/publications/statistical/attribution-dataset-gp-registered-populations/attribution-data-set-gp-registered-populations-scaled-to-ons-population-estimates-2011 (NHS Digital, 2012).

28. Cattaneo, M. D., Keele, L. 和 Titiunik, R. 醫療應用中的回歸不連續設計指南。統計醫學 42, 4484–4513（2023 年）。
    Cattaneo, M. D., Keele, L. & Titiunik, R. A guide to regression discontinuity designs in medical applications. Stat. Med. 42, 4484–4513 (2023).

    Article MathSciNet PubMed Google Scholar 

29. Cattaneo, M. D., Idrobo, N. & Titiunik, R. 實用的回歸不連續設計入門：基礎（劍橋大學出版社，2020 年）。
    Cattaneo, M. D., Idrobo, N. & Titiunik, R. A Practical Introduction to Regression Discontinuity Designs: Foundations (Cambridge Univ. Press, 2020).

30. Harbecke, R., Cohen, J. I. & Oxman, M. N. 帶狀皰疹疫苗。感染病學雜誌 224, S429–S442 (2021 年)。
    Harbecke, R., Cohen, J. I. & Oxman, M. N. Herpes zoster vaccines. J. Infect. Dis. 224, S429–S442 (2021).

    Article CAS PubMed PubMed Central Google Scholar 

31. 2024 年阿茲海默症事實與數據。阿茲海默症與癡呆症 20, 3708–3821（2024 年）。
    2024 Alzheimer’s disease facts and figures. Alzheimers Dement. 20, 3708–3821 (2024).

32. Walters, K. 等人。預測初級保健中的癡呆風險：使用常規收集數據開發和驗證癡呆風險評分。BMC Med. 14, 6 (2016)。
    Walters, K. et al. Predicting dementia risk in primary care: development and validation of the Dementia Risk Score using routinely collected data. BMC Med. 14, 6 (2016).

    Article CAS PubMed PubMed Central Google Scholar 

33. 2019 年全球疾病負擔項目。GBD Compare vizhub.healthdata.org/gbd-compare（健康指標與評估研究所，2020 年）。
    Global Burden of Disease Project 2019. GBD Compare vizhub.healthdata.org/gbd-compare (Institute for Health Metrics and Evaluation, 2020).

34. Charlson, M. E., Pompei, P., Ales, K. L. 和 MacKenzie, C. R. 一種在縱向研究中分類預後合併症的新方法：開發和驗證。J. Chronic Dis. 40, 373–383 (1987)。
    Charlson, M. E., Pompei, P., Ales, K. L. & MacKenzie, C. R. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J. Chronic Dis. 40, 373–383 (1987).

    Article CAS PubMed Google Scholar 

35. Oxman, M. N. 帶狀皰疹疫苗：當前狀況和未來前景。Clin. Infect. Dis. 51, 197–213 (2010)。
    Oxman, M. N. Zoster vaccine: current status and future prospects. Clin. Infect. Dis. 51, 197–213 (2010).

    Article PubMed Google Scholar 

36. Oxman, M. N. 等人。預防帶狀皰疹及老年人後帶狀皰疹神經痛的疫苗。新英格蘭醫學雜誌 352, 2271–2284 (2005)。
    Oxman, M. N. et al. A vaccine to prevent herpes zoster and postherpetic neuralgia in older adults. N. Engl. J. Med. 352, 2271–2284 (2005).

    Article CAS PubMed Google Scholar 

37. Campling, J. 等人。針對有增加肺炎球菌疾病風險的成年人進行肺炎球菌疫苗接種的證據回顧：風險組別定義及英國疫苗接種覆蓋率的優化。專家疫苗回顧 22, 785–800 (2023)。
    Campling, J. et al. A review of evidence for pneumococcal vaccination in adults at increased risk of pneumococcal disease: risk group definitions and optimization of vaccination coverage in the United Kingdom. Expert Rev. Vaccines 22, 785–800 (2023).

    Article CAS PubMed Google Scholar 

38. Yun, H. 等人。自身免疫及炎症性疾病中帶狀皰疹的風險：對疫苗接種的啟示。關節炎風濕病學 68, 2328–2337 (2016)。
    Yun, H. et al. Risk of herpes zoster in autoimmune and inflammatory diseases: implications for vaccination. Arthritis Rheumatol. 68, 2328–2337 (2016).

    Article CAS PubMed PubMed Central Google Scholar 

39. Marra, F., Parhar, K., Huang, B. & Vadlamudi, N. 帶狀皰疹感染的風險因素：一項元分析。開放論壇感染病 7, ofaa005 (2020)。
    Marra, F., Parhar, K., Huang, B. & Vadlamudi, N. Risk factors for herpes zoster infection: a meta-analysis. Open Forum Infect. Dis. 7, ofaa005 (2020).

    Article PubMed PubMed Central Google Scholar 

40. Joesoef, R. M., Harpaz, R., Leung, J. & Bialek, S. R. 慢性疾病作為帶狀皰疹的風險因素。梅奧診所程序 87, 961–967 (2012)。
    Joesoef, R. M., Harpaz, R., Leung, J. & Bialek, S. R. Chronic medical conditions as risk factors for herpes zoster. Mayo Clin. Proc. 87, 961–967 (2012).

    Article PubMed PubMed Central Google Scholar 

41. Kawai, K. & Yawn, B. P. 帶狀皰疹的風險因素：系統性回顧和元分析。梅奧診所程序 92, 1806–1821 (2017)。
    Kawai, K. & Yawn, B. P. Risk factors for herpes zoster: a systematic review and meta-analysis. Mayo Clin. Proc. 92, 1806–1821 (2017).

    Article PubMed Google Scholar 

42. McGonagle, D. & McDermott, M. F. 提議的免疫疾病分類。PLoS 醫學 3, e297 (2006)。
    McGonagle, D. & McDermott, M. F. A proposed classification of the immunological diseases. PLoS Med. 3, e297 (2006).

    Article PubMed PubMed Central Google Scholar 

43. Gandhi, N. A. 等。針對疾病中 2 型炎症的關鍵近端驅動因素。自然評論：藥物發現 15, 35–50 (2016)。
    Gandhi, N. A. et al. Targeting key proximal drivers of type 2 inflammation in disease. Nat. Rev. Drug Discov. 15, 35–50 (2016).

    Article CAS PubMed Google Scholar 

44. 帶狀皰疹疫苗覆蓋率（英格蘭）：2022 至 2023 財政年度第一季度報告 www.gov.uk/government/publications/ herpes-zoster-shingles- immunisation-programme-2022-to-2023- evaluation-reports/shingles-vaccine- coverage-england-report-for-quarter-1- of-the-financial-year-2022-to-2023 (英國健康安全局，2024 年)。
    Shingles Vaccine Coverage (England): Report for Quarter 1 of the Financial Year 2022 to 2023 www.gov.uk/government/publications/ herpes-zoster-shingles- immunisation-programme-2022-to-2023- evaluation-reports/shingles-vaccine- coverage-england-report-for-quarter-1- of-the-financial-year-2022-to-2023 (UK Health Security Agency, 2024).

45. 人口和家庭估算，英格蘭和威爾士：2021 年人口普查，未四捨五入數據 www.ons.gov.uk/peoplepopulationandcommunity/ populationandmigration/populationestimates/ bulletins/ populationandhouseholdestimatesenglandandwales/ census2021unroundeddata (國家統計局，2022 年)。
    Population and Household estimates, England and Wales: Census 2021, Unrounded Data www.ons.gov.uk/peoplepopulationandcommunity/ populationandmigration/populationestimates/ bulletins/ populationandhouseholdestimatesenglandandwales/ census2021unroundeddata (ONS, 2022).

46. Fischinger, S., Boudreau, C. M., Butler, A. L., Streeck, H. & Alter, G. 疫苗引起的體液免疫中的性別差異。Semin. Immunopathol. 41, 239–249 (2019 年)。
    Fischinger, S., Boudreau, C. M., Butler, A. L., Streeck, H. & Alter, G. Sex differences in vaccine-induced humoral immunity. Semin. Immunopathol. 41, 239–249 (2019).

    Article CAS PubMed Google Scholar 

47. Ferretti, M. T. 等。阿茲海默症中的性別差異：當前挑戰及對臨床實踐的影響：歐洲神經學學院癡呆症和認知障礙小組的立場文件。歐洲神經學雜誌 27, 928–943 (2020 年)。
    Ferretti, M. T. et al. Sex and gender differences in Alzheimer’s disease: current challenges and implications for clinical practice: position paper of the Dementia and Cognitive Disorders Panel of the European Academy of Neurology. Eur. J. Neurol. 27, 928–943 (2020).

    Article CAS PubMed Google Scholar 

48. Silver, B. 等人。水痘帶狀疱疹病毒血管病：一種可治療的快速進展性多梗塞癡呆，持續 2 年後出現。神經科學雜誌 323, 245–247 (2012)。
    Silver, B. et al. Varicella zoster virus vasculopathy: a treatable form of rapidly progressive multi-infarct dementia after 2 years’ duration. J. Neurol. Sci. 323, 245–247 (2012).

    Article PubMed PubMed Central Google Scholar 

49. Gilden, D. 等人。成功的抗病毒治療，經過 6 年的慢性進行性神經疾病，歸因於水痘帶狀疱疹病毒腦部感染。神經科學雜誌 368, 240–242 (2016)。
    Gilden, D. et al. Successful antiviral treatment after 6years of chronic progressive neurological disease attributed to VZV brain infection. J. Neurol. Sci. 368, 240–242 (2016).

    Article PubMed PubMed Central Google Scholar 

50. Bubak, A. N. et al. 針對水痘帶狀皰疹病毒感染的腦血管外膜成纖維細胞進行的定向 RNA 測序顯示，淀粉樣蛋白可能與水痘帶狀皰疹病毒血管病變有關。神經學. 神經免疫學. 神經炎症. 9, e1103 (2022)。
    Bubak, A. N. et al. Targeted RNA sequencing of VZV-infected brain vascular adventitial fibroblasts indicates that amyloid may be involved in VZV vasculopathy. Neurol. Neuroimmunol. Neuroinflamm. 9, e1103 (2022).

    Article PubMed Google Scholar 

51. Nagel, M. A. & Bubak, A. N. 水痘帶狀皰疹病毒血管病。感染疾病雜誌 218, S107–S112 (2018)。
    Nagel, M. A. & Bubak, A. N. Varicella zoster virus vasculopathy. J. Infect. Dis. 218, S107–S112 (2018).

    Article PubMed PubMed Central Google Scholar 

52. Nagel, M. A. 等。水痘帶狀皰疹病毒血管病：臨床、腦脊液、影像學和病毒學特徵。神經學 70, 853–860 (2008)。
    Nagel, M. A. et al. The varicella zoster virus vasculopathies: clinical, CSF, imaging, and virologic features. Neurology 70, 853–860 (2008).

    Article CAS PubMed Google Scholar 

53. Nagel, M. A. 等。水痘帶狀皰疹病毒血管病：病毒感染動脈的分析。神經學 77, 364–370 (2011)。
    Nagel, M. A. et al. Varicella zoster virus vasculopathy: analysis of virus-infected arteries. Neurology 77, 364–370 (2011).

    Article CAS PubMed PubMed Central Google Scholar 

54. Nagel, M. A. 等。水痘帶狀皰疹病毒血管病：病毒感染動脈的免疫特徵。神經學 80, 62–68 (2013)。
    Nagel, M. A. et al. Varicella-zoster virus vasculopathy: immune characteristics of virus-infected arteries. Neurology 80, 62–68 (2013).

    Article CAS PubMed PubMed Central Google Scholar 

55. Attems, J. & Jellinger, K. A. 血管疾病與阿茲海默症之間的重疊——來自病理學的教訓。BMC Med. 12, 206 (2014).
    Attems, J. & Jellinger, K. A. The overlap between vascular disease and Alzheimer’s disease-lessons from pathology. BMC Med. 12, 206 (2014).

    Article PubMed PubMed Central Google Scholar 

56. Boyle, P. A. 等人。社區基礎老年人中的腦淀粉樣血管病與認知結果。神經學 85, 1930–1936 (2015).
    Boyle, P. A. et al. Cerebral amyloid angiopathy and cognitive outcomes in community-based older persons. Neurology 85, 1930–1936 (2015).

    Article CAS PubMed PubMed Central Google Scholar 

57. Cairns, D. M., Itzhaki, R. F. & Kaplan, D. L. 水痘帶狀疱疹病毒在阿茲海默症中的潛在參與，通過重新激活潛伏的單純疱疹病毒 1 型。阿茲海默病雜誌 88, 1189–1200 (2022)。
    Cairns, D. M., Itzhaki, R. F. & Kaplan, D. L. Potential involvement of varicella zoster virus in Alzheimer’s disease via reactivation of quiescent herpes simplex virus type 1. J. Alzheimers Dis. 88, 1189–1200 (2022).

    Article CAS PubMed Google Scholar 

58. Bukhbinder, A. S., Ling, Y., Harris, K., Jiang, X. & Schulz, P. E. 疫苗接種是否影響阿茲海默症的發展？人類疫苗與免疫治療 19, 2216625 (2023)。
    Bukhbinder, A. S., Ling, Y., Harris, K., Jiang, X. & Schulz, P. E. Do vaccinations influence the development of Alzheimer disease? Hum. Vaccin. Immunother. 19, 2216625 (2023).

    Article PubMed PubMed Central Google Scholar 

59. Anderson, M. 等。英國：健康系統評估。健康系統過渡 24, 1–194 (2022)。
    Anderson, M. et al. United Kingdom: health system review. Health Syst. Transit. 24, 1–194 (2022).

    PubMed Google Scholar 

60. Ford, D. V. 等人。SAIL 數據庫：建立一個國家架構以進行電子健康研究和評估。BMC 健康服務研究 9, 157 (2009)。
    Ford, D. V. et al. The SAIL Databank: building a national architecture for e-health research and evaluation. BMC Health Serv. Res. 9, 157 (2009).

    Article PubMed PubMed Central Google Scholar 

61. Jones, K. H. 等人。安全匿名信息鏈接 (SAIL) 閘道的案例研究：一個保護隱私的健康相關研究和評估的遠程訪問系統。生物醫學信息學雜誌 50, 196–204 (2014)。
    Jones, K. H. et al. A case study of the Secure Anonymous Information Linkage (SAIL) Gateway: a privacy-protecting remote access system for health-related research and evaluation. J. Biomed. Inform. 50, 196–204 (2014).

62. Lyons, R. A. 等人。SAIL 數據庫：鏈接多個健康和社會護理數據集。BMC 醫學信息與決策制定 9, 3 (2009)。
    Lyons, R. A. et al. The SAIL databank: linking multiple health and social care datasets. BMC Med. Inform. Decis. Mak. 9, 3 (2009).

63. Rodgers, S. E., Demmler J., Dsilva R. 和 Lyons R. 在使用基於居住地的環境和人口統計數據時保護健康數據隱私。健康地點 18, 209–217 (2012)。
    Rodgers, S. E., Demmler J., Dsilva R. & Lyons R. Protecting health data privacy while using residence-based environment and demographic data. Health Place 18, 209–217 (2012).

64. Rodgers, S. E. 等人。住宅匿名連結領域（RALFs）：一種新穎的信息基礎設施，用於研究環境與個人健康之間的互動。《公共衛生雜誌》31, 582–588 (2009)。
    Rodgers, S. E. et al. Residential Anonymous Linking Fields (RALFs): a novel information infrastructure to study the interaction between the environment and individuals’ health. J. Public Health 31, 582–588 (2009).

65. 威爾士人口服務數據集（WDSD）網頁：www.healthdatagateway.org/dataset/cea328df-abe5-48fb-8bcb-c0a5b6377446#（數字健康與護理威爾士，2022）。
    Welsh Demographic Service Dataset (WDSD) web.www.healthdatagateway.org/dataset/cea328df-abe5-48fb-8bcb-c0a5b6377446# (Digital Health and Care Wales, 2022).

66. 威爾士縱向全科醫療數據集（WLGP）—威爾士初級護理網頁：www.healthdatagateway.org/dataset/33fc3ffd-aa4c-4a16-a32f-0c900aaea3d2（SAIL 數據庫，2023）。
    Welsh Longitudinal General Practice Dataset (WLGP)—Welsh Primary Care web.www.healthdatagateway.org/dataset/33fc3ffd-aa4c-4a16-a32f-0c900aaea3d2 (SAIL Databank, 2023).

67. Benson, T. 《Read 代碼的歷史：2011 年詹姆斯·Read 紀念講座的開幕演講》。信息初級護理 19, 173–182 (2011)。
    Benson, T. The history of the Read Codes: the inaugural James Read Memorial Lecture 2011. Inform. Prim. Care 19, 173–182 (2011).

    PubMed Google Scholar 

68. 威爾士患者事件數據庫 dhcw.nhs.wales/information-services/health-intelligence/pedw-data-online/（數字健康與護理威爾士，2020 年）。
    Patient Episode Database for Wales dhcw.nhs.wales/information-services/health-intelligence/pedw-data-online/ (Digital Health and Care Wales, 2020).

69. OPCS 介入和程序分類 www.datadictionary.nhs.uk/supporting_information/opcs_classification_of_interventions_and_procedures.html（英國國民健康服務，2023 年）。
    OPCS Classification of Interventions and Procedures www.datadictionary.nhs.uk/supporting_information/opcs_classification_of_interventions_and_procedures.html (NHS England, 2023).

70. ICD-10：國際疾病和相關健康問題統計分類：第十版 apps.who.int/iris/handle/10665/42980（世界衛生組織，2004 年）。
    ICD-10: International Statistical Classification of Diseases and Related Health Problems: Tenth Revision apps.who.int/iris/handle/10665/42980 (WHO, 2004).

71. 威爾士門診數據庫 (OPDW) web.www.healthdatagateway.org/dataset/d331159b-b286-4ab9-8b36-db39123ec229 (數字健康與護理威爾士, 2022)。
    Outpatient Database for Wales (OPDW) web.www.healthdatagateway.org/dataset/d331159b-b286-4ab9-8b36-db39123ec229 (Digital Health and Care Wales, 2022).

72. 威爾士癌症情報與監測單位 (WCISU) phw.nhs.wales/services-and-teams/welsh-cancer-intelligence-and-surveillance-unit-wcisu/ (威爾士公共衛生, 2023)。
    Welsh Cancer Intelligence and Surveillance Unit (WCISU) phw.nhs.wales/services-and-teams/welsh-cancer-intelligence-and-surveillance-unit-wcisu/ (Public Health Wales, 2023).

73. 年度地區死亡提取 (ADDE) web.www.healthdatagateway.org/dataset/15cf4241-abad-4dcc-95b0-8cd7c02be999 (數字健康與護理威爾士, 2023)。
    Annual District Death Extract (ADDE) web.www.healthdatagateway.org/dataset/15cf4241-abad-4dcc-95b0-8cd7c02be999 (Digital Health and Care Wales, 2023).

74. CT1303_2011 人口普查 www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationestimates/adhocs/2089ct13032011census (英國國家統計局, 2024)。
    CT1303_2011 Census www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationestimates/adhocs/2089ct13032011census (ONS, 2024).

75. Bricout, H., Haugh, M., Olatunde, O. & Prieto, R. G. 歐洲帶狀皰疹相關死亡率：系統性回顧。BMC 公共衛生 15, 466 (2015)。
    Bricout, H., Haugh, M., Olatunde, O. & Prieto, R. G. Herpes zoster-associated mortality in Europe: a systematic review. BMC Public Health 15, 466 (2015).

    Article PubMed PubMed Central Google Scholar 

76. Stoltenberg, E. A. 右截尾生存數據的回歸不連續設計。預印本於 arxiv.org/abs/2210.02548 (2022)。
    Stoltenberg, E. A. Regression discontinuity design with right-censored survival data. Preprint at arxiv.org/abs/2210.02548 (2022).

77. 阿茲海默症協會。2022 年阿茲海默症疾病事實與數據。阿茲海默症與癡呆症 18, 700–789 (2022)。
    Alzheimer’s Association. 2022 Alzheimer’s disease facts and figures. Alzheimers Dement. 18, 700–789 (2022).

    Article Google Scholar 

78. Boyle, P. A. et al. 與年齡相關的神經病理學所導致的阿茲海默症的可歸因風險。Ann. Neurol. 85, 114–124 (2019)。
    Boyle, P. A. et al. Attributable risk of Alzheimer’s dementia attributed to age-related neuropathologies. Ann. Neurol. 85, 114–124 (2019).

    Article CAS PubMed Google Scholar 

79. Boyle, P. A. 等人。特定於個體的神經病理學對老年認知喪失的貢獻。《神經學年鑑》83, 74–83 (2018)。
    Boyle, P. A. et al. Person-specific contribution of neuropathologies to cognitive loss in old age. Ann. Neurol. 83, 74–83 (2018).

    Article CAS PubMed PubMed Central Google Scholar 

80. Villamizar-Villegas, M., Pinzon-Puerto, F. A. & Ruiz-Sanchez, M. A. 回歸不連續設計的綜合歷史：對過去 60 年的實證調查。《經濟調查期刊》36, 1130–1178 (2022)。
    Villamizar-Villegas, M., Pinzon-Puerto, F. A. & Ruiz-Sanchez, M. A. A comprehensive history of regression discontinuity designs: an empirical survey of the last 60 years. J. Econ. Surv. 36, 1130–1178 (2022).

81. Gelman, A. & Imbens, G. 為什麼在回歸不連續設計中不應使用高階多項式。商業經濟統計雜誌 https://doi.org/10.1080/07350015.2017.1366909 (2017)。
    Gelman, A. & Imbens, G. Why high-order polynomials should not be used in regression discontinuity designs. J. Bus. Econ. Stat. https://doi.org/10.1080/07350015.2017.1366909 (2017).

82. Imbens, G. & Kalyanaraman, K. 回歸不連續估計器的最佳帶寬選擇。《經濟研究評論》79, 933–959 (2012)。
    Imbens, G. & Kalyanaraman, K. Optimal bandwidth choice for the regression discontinuity estimator. Rev. Econ. Stud. 79, 933–959 (2012).

    Article MathSciNet Google Scholar 

83. Calonico, S., Cattaneo, M. D. & Titiunik, R. 回歸不連續設計的穩健非參數置信區間。《計量經濟學》82, 2295–2326 (2014)。
    Calonico, S., Cattaneo, M. D. & Titiunik, R. Robust nonparametric confidence intervals for regression-discontinuity designs. Econometrica 82, 2295–2326 (2014).

    Article MathSciNet Google Scholar 

84. Imbens, G. W. & Lemieux, T. 回歸不連續設計：實踐指南。經濟計量學雜誌 142, 615–635 (2008)。
    Imbens, G. W. & Lemieux, T. Regression discontinuity designs: a guide to practice. J. Econometr. 142, 615–635 (2008).

    Article MathSciNet Google Scholar 

85. Bor, J., Moscoe, E., Mutevedzi, P., Newell, M. L. & Barnighausen, T. 回歸不連續設計在流行病學中的應用：無隨機試驗的因果推斷。流行病學 25, 729–737 (2014)。
    Bor, J., Moscoe, E., Mutevedzi, P., Newell, M. L. & Barnighausen, T. Regression discontinuity designs in epidemiology: causal inference without randomized trials. Epidemiology 25, 729–737 (2014).

    Article PubMed PubMed Central Google Scholar 

86. Oehlert, G. W. 關於德爾塔方法的說明。美國統計學雜誌 46, 27–29 (1992)。
    Oehlert, G. W. A note on the delta method. Am. Stat. 46, 27–29 (1992).

    Article MathSciNet Google Scholar 

87. Kolesár, M. & Rothe, C. 在具有離散運行變量的回歸不連續設計中的推斷。美國經濟學評論 108, 2277–2304 (2018)。
    Kolesár, M. & Rothe, C. Inference in regression discontinuity designs with a discrete running variable. Am. Econ. Rev. 108, 2277–2304 (2018).

    Article Google Scholar 

88. Armstrong, T. B. & Kolesár, M. 在非參數回歸中簡單且誠實的置信區間。定量經濟學 11, 1–39 (2020)。
    Armstrong, T. B. & Kolesár, M. Simple and honest confidence intervals in nonparametric regression. Quant. Econ. 11, 1–39 (2020).

    Article MathSciNet Google Scholar 

89. VanderWeele, T. J. 中介分析：實務指南。公共衛生年評 37, 17–32 (2016)。
    VanderWeele, T. J. Mediation analysis: a practitioner’s guide. Ann. Rev. Publ. Health 37, 17–32 (2016).

    Article Google Scholar 

90. Benjamini, Y. & Hochberg, Y. 控制虛假發現率：一種實用且強大的多重測試方法。J. R. Stat. Soc. B 57, 289–300 (1995)。
    Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J. R. Stat. Soc. B 57, 289–300 (1995).

    Article MathSciNet Google Scholar 

91. Conrad, N. et al. 自體免疫疾病與心血管風險：在 2200 萬英國人中對 19 種自體免疫疾病和 12 種心血管疾病的基於人群的研究。柳葉刀 400, 733–743 (2022)。
    Conrad, N. et al. Autoimmune diseases and cardiovascular risk: a population-based study on 19 autoimmune diseases and 12 cardiovascular diseases in 22 million individuals in the UK. Lancet 400, 733–743 (2022).

    Article PubMed Google Scholar 

92. Mukherjee, M., Wyatt, J. C., Simpson, C. R. & Sheikh, A. 在初級保健電子健康記錄中使用過敏代碼：蘇格蘭的全國評估。過敏 71, 1594–1602 (2016)。
    Mukherjee, M., Wyatt, J. C., Simpson, C. R. & Sheikh, A. Usage of allergy codes in primary care electronic health records: a national evaluation in Scotland. Allergy 71, 1594–1602 (2016).

    Article CAS PubMed Google Scholar 

93. Xie, M. 等。關於帶狀皰疹疫苗接種對癡呆症影響的自然實驗。預印本於 https://osf.io/cfnr6/?view_only=d3774e4fda2649e2b2031431b1234874 (2025)。
    Xie, M. et al. A natural experiment on the effect of herpes zoster vaccination on dementia. Preprint at https://osf.io/cfnr6/?view_only=d3774e4fda2649e2b2031431b1234874 (2025).