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

Data source and study period

The study cohort was obtained from Optum’s de-identified Clinformatics® Data Mart Database (CDM). The claims database records information from different sources in the United States, such as medical, pharmaceutical, and administrative claims, as well as laboratory test results. The database includes patients who have both medical and prescription drug coverage through private insurance or Medicare Advantage with Part D. Mortality information from hospital discharge claims and the Social Security Administration Death Master file is also available in the CDM. All data are verified, adjudicated, adjusted, and de-identified before inclusion in the CDM.

For our study, the CDM includes the years 2009 through 2019. With the exception of three sub-analyses (as discussed in the Analysis Overview section below), all analyses were performed using a look-back period of September 1, 2009 to August 31, 2011 and a follow-up period of September 1, 2011 to August 31, 2019.

Cohort selection

With the definition of the look-back period and the follow-up period, we implemented inclusion and exclusion criteria to build a cohort for analyzing the effects of the targeted vaccines (Fig. 1).

Fig. 1

Flowchart of Sampling Methodology. The three main analyses using Tdap/Td, HZ, and pneumococcal vaccinations are shown. AD, Alzheimer’s disease; CDM, Optum’s de-identified Clinformatics® Data Mart Database; HZ, Herpes zoster; ICD, International Classification of Diseases; Tdap/Td, Tetanus toxoid, reduced diphtheria toxoid, and acellular pertussis/Tetanus toxoid, and reduced diphtheria toxoid. Figure adapted from Bukhbinder et al. [38]. Reprinted from Journal of Alzheimer’s Disease, vol. 88, no. 3, Bukhbinder AS, Ling Y, Hasan O, Jiang X, Kim Y, Phelps KN, Schmandt RE, Amran A, Coburn R, Ramesh S, Xiao Q, Schulz PE, Risk of Alzheimer’s disease following influenza vaccination: a claims-based cohort study using propensity score matching, pp. 1061-1074, 2022, with permission from IOS Press. The publication is available at IOS Press through http://dx.doi.org/10.3233/JAD-220361.

We included patients who were at least 65 years old at the start of the follow-up period. Patients were included if they had at least one record in the look-back period and had at least two records during the follow-up. If patients had 1) a recorded diagnosis of dementia, mild cognitive impairment, or encephalopathy, or 2) were prescribed any medication primarily indicated for AD (i.e., donepezil, galantamine, rivastigmine, or memantine) during the look-back period, they were excluded from the cohort.

Exposure measurement

Vaccinations were counted if they were received on or after the index date (i.e., the first day of the follow-up period) and before the following occurred: 1) AD onset, 2) death, or 3) the end of the follow-up period. We investigated three kinds of vaccination in this study: Tdap/Td, HZ, and pneumococcal vaccines. To identify vaccinations, we queried the database for their brand names and generic names as found in Supplementary Table 1. For the Tdap/Td vaccine sample, we excluded vaccines not indicated for patients 65 years and older (i.e., DTaP). For the HZ vaccines, only the two brands of vaccines approved by the FDA for use in the U.S. were included: Zostavax and Shingrix. And for the pneumococcal vaccines, we included PCV13 and PPSV23, while excluding Pneumococcal 7-val vaccines as they are only used for pediatric patients [26].

Outcome measurement

The procedure and rationale for outcome measurement is the same as what was used in our recent study of incident AD risk following influenza vaccination [38]. We identified patients as having AD if they met any of the following three criteria in any 12-month window during the follow-up period: 1) two or more diagnoses of AD in their records, 2) one or more diagnoses of AD and one or more prescription records for AD-related medications, or 3) two or more prescription records for AD-related medications. Patients who only have one record of an AD diagnosis or AD-related prescription were removed from the cohort. The ICD codes and medications used for identifying AD are located in Supplementary Table 1. A systematic review of validation studies for AD and ADRD in administrative datasets provide support for our inclusion and exclusion criteria for the outcome measurement [39]. The authors found that the positive predictive value (PPV) of a patient having dementia increased from 68% to 94% if two or more diagnosis codes were utilized instead of just one. Further, they found that the PPV was 97% when using AD medication codes to identify patients with AD. Lastly, we elected to make use of nonspecific dementia codes, as well as AD specific codes, in identifying AD patients. This is because, although 60–70% of dementia cases among older adults are secondary to AD, nonspecific dementia codes (e.g., senile dementia) are significantly more common than codes for specific dementia subtypes (e.g., AD, vascular dementia) in administrative claims data [40, 41]. For example, a study of Medicare beneficiaries found that 46.1% of patients only had a code for dementia not otherwise specified, 4.5% of patients only had a code for AD, and 29% of patients had codes for both dementia not otherwise specified andfor AD [40].

Covariate measurement

Similar to our previous research on influenza vaccination and AD risk [38], and to another study on influenza vaccination and dementia in a Veterans Affairs cohort [42], we included covariates for patient demographics, comorbidities, medication use, and the number of healthcare encounters and routine “well visit” examinations (as proxies for healthcare utilization rate). For this analysis, we also included information pertaining to receipt of routine vaccinations, including those against tetanus, diphtheria, with or without pertussis; HZ; pneumococcus; and influenza. Importantly, the vaccine(s) used in the exposure definition (see “Analysis Overview” below) for a given analysis was not included as a covariate in that analysis; for example, in the analysis comparing persons who received either Tdap or Td with those who received neither during follow-up, Tdap and Td vaccinations during the look-back period were not included as a covariate in the propensity score model. A detailed list of the covariates and their definitions is provided in Supplementary Table 1. For all covariates except age, the last measurement recorded in the look-back period was used as the baseline covariate value; age on the first day of the follow-up period was used as the baseline covariate value.

Estimating ATT using propensity score matching

We estimate the average treatment effect on the treated (ATT) of the three vaccination groups on AD risk using propensity score matched (PSM) (Fig. 2). We utilized PSM to minimize selection bias from unbalanced confounders between the vaccinated and unvaccinated groups. The propensity scores were estimated by fitting a logistic regression model with all the baseline characteristics measured during the look-back period to predict the probability of vaccination. For non-static variables (e.g., BMI), the last measurement in the look-back period (i.e., the one closest to the start of follow-up) was used. We assumed that receiving one kind of vaccine would lead to a higher probability of receiving other kinds of adult vaccines, and therefore, we included other routine vaccines as covariates (see “Covariate Measurement” above). Patients with unknown sex, geographic region, or race were excluded from this analysis. Once we estimated the propensity scores using logistic regression, a one-to-one nearest neighbor matching with a caliper width of 0.2 standard deviations of the logit of the propensity score and without replacement was used to match each patient that met target vaccine group criterion with a patient in the unvaccinated group [43]. To evaluate the balance between vaccinated and unvaccinated groups after matching, we calculated the standardized mean difference (SMD) for each covariate before and after matching. An adequate balance between the groups was defined as an SMD≤0.10 [44].

Fig. 2

Overview of Cohort Selection and Propensity Score Matching. AD, Alzheimer’s disease; ICD, International Classification of Diseases. Figure adapted from Bukhbinder et al. [38]. Reprinted from Journal of Alzheimer’s Disease, vol. 88, no. 3, Bukhbinder AS, Ling Y, Hasan O, Jiang X, Kim Y, Phelps KN, Schmandt RE, Amran A, Coburn R, Ramesh S, Xiao Q, Schulz PE, Risk of Alzheimer’s disease following influenza vaccination: a claims-based cohort study using propensity score matching, pp. 1061-1074, 2022, with permission from IOS Press. The publication is available at IOS Press through http://dx.doi.org/10.3233/JAD-220361.

Analysis overview

We performed three main analyses and then separate sub-analyses for each of the three vaccines under study. In these analyses, we created vaccinated and unvaccinated balanced cohorts by PSM and estimated ATT in order to evaluate for heterogeneity in the effect size on the risk of AD among the vaccines targeting the same pathogenic species. Each analysis performed had a different unvaccinated cohort. There were thirteen analyses performed in total.

For the Tdap/Td vaccine, the main analysis was performed on patients who were vaccinated with either Tdap and Td as the exposed group; compared with Tdap/Td- unvaccinated patients in an unexposed cohort. We included four other sub-analyses: patients who received 1) at least one Tdap, Td, or TT vaccine; 2) at least one Tdap vaccine; 3) at least one Td vaccine; and, 4) at least one TT vaccine. The comparison group for the Tdap/Td analyses consisted of adults who received no Tdap, Td, or TT vaccines.

With regard to HZ vaccines, the main analysis included patients who received at least one Zostavax or at least one Shingrix vaccine compared of those who received neither. The four sub-analyses included patients who 1) were fully vaccinated using the Shingrix vaccine (completed two doses of the vaccine); 2) received at least one Zostavax vaccine and were fully vaccinated using the Shingrix vaccine; 3) received at least one Shingrix vaccine but no Zostavax vaccine; and, 4) received at least one Zostavax vaccine but no Shingrix vaccine. The comparison group for all of the HZ vaccine analyses consisted of adults who received neither Shingrix nor Zostavax.

For the pneumococcal vaccines, the main analysis included patients who received at least one PCV-13 vaccine or PPSV-23 vaccine compared of those who received neither. The two sub-analyses were for patients who received 1) at least one PCV-13 vaccine, but no PPSV-23 vaccine; and 2) at least one PPSV-23 vaccine, but no PCV-13 vaccine. The comparison group for all of the pneumococcal vaccine analyses consisted of adults who received neither PCV-13 nor PPSV-23.

The look-back and follow-up periods were 2009–2011 and 2011–2019 for most of the analyses, with three exceptions that were necessary to account for the year in which two of the vaccines (Shingrix and PCV-13) were added to the CDC’s routine immunization schedule for older adults. As discussed earlier, Shingrix was first approved and recommended for use in 2017 [15, 21]. Hence, for the sub-analyses in which the treatment (vaccinated) group consisted of patients who received at least one Shingrix vaccination but no Zostavax vaccination, or who received the full Shingrix series (two doses) but no Zostavax vaccines, we set the look-back period to 2009–2017 and the follow-up period to 2017–2019. Similarly, because the PCV-13 vaccine was first recommended for older adults in 2014, the sub-analysis of patients who received at least one PCV-13 vaccination but no PPSV-23 used a look-back period spanning 2009–2014 and a follow-up period spanning 2014–2019 [25, 27, 28].

For all of the analyses, we computed relative risk (RR), absolute risk reduction (ARR), and the corresponding 95% confidence intervals (CIs). When constructing the 95% CI for the point estimates, given that the study cohort is propensity-score-matched cohort, we used a method that accounts for the pairwise dependence between matched samples [45, 46]. E-values for point estimates were calculated to assess how strongly an unmeasured confounder would need to be associated with both the probability of vaccination and the probability of AD, while controlling for the covariates in our analyses, in order to render the results statistically insignificant at a significance level of α=0.05. For example, if the E-value for the RR of an analysis is 4, then an unmeasured confounder would need to have a RR of≥4 (while controlling for the same covariates) with both the exposure (vaccination) and with the outcome (incident AD) for the result to become statistically insignificant. PSM was conducted with Python 3.7.7 and CausalML package v0.11.1 [47].

Sensitivity analysis: Controlling for healthy adherer bias

To investigate the influence of healthy adherer bias, we applied the eligibility criteria described above but then selected a subset of patients who filled at least one statin (i.e., HMG-CoA reductase inhibitor) prescription in the first half of the look-back period (2009–2010) and whose proportion of days covered (PDC) for statin therapy during the second half of the look-back period (2010–2011) was≥80%. The remainder of the primary analysis (i.e., ATT estimation using propensity-score matching) was repeated using this subset of statin adherers.

Ethics approval

This study was reviewed by the UTHealth Institutional Review Board, the Committee for the Protection of Human Subjects (CPHS), which deemed this study “non-human subjects research” because the study uses de-identified retrospective claims data. Therefore, the study was approved with a waiver of the HIPAA authorization and waiver of informed consent.

Baseline characteristics

In total, 1,651,991 patients were identified after applying inclusion and exclusion criteria (Fig. 1). Prior to matching, 122,777 patients received vaccinations against tetanus and diphtheria, with or without pertussis, during the follow-up period; 212,417 received vaccinations against herpes zoster; and 286,504 received vaccines against pneumococcus. Summary of baseline characteristics before and after PSM for Tdap/Td is shown in Table 1, and for HZ and pneumococcal is shown in Supplementary Table 2A and 2B. In the analyses, the SMDs of all covariates were less than or equal to 0.1 after PSM, which indicates adequate covariate balance between the matched groups.

ATT estimation

The frequency of AD among patients who were vaccinated and unvaccinated after PSM for our main analyses and sub-analyses are shown in Table 2. In the main analyses, for the Tdap/Td vaccine, 7.2% (n = 8,370) of the vaccinated patients and 10.2% (n = 11,857) of the unvaccinated patients developed AD during the 8-year follow-up period. For the HZ vaccine, 8.1% (n = 16,106) of the vaccinated patients and 10.7% (n = 21,417) of the unvaccinated patients developed AD during the 8-year follow-up period. And for the pneumococcal vaccine, 7.92% (n = 20,583) of the vaccinated patients and 10.9% (n = 28,558) of the unvaccinated patients developed AD during the 8-year follow-up period. The estimated RR, ARR, number needed to treat (NNT) and E-values for the thirteen different analyses are shown in Table 3. All three main analyses showed statistically significant results: Tdap/Td vaccination (RR: 0.70; 95% CI: 0.68–0.72), HZ vaccination (RR: 0.75; 95% CI: 0.73–0.76), and pneumococcal vaccination (RR: 0.73; 95% CI: 0.71–0.74). There were also statistically significant results in several sub-analyses including: 1) at least one dose of Shingrix (excluding any Zostavax vaccinations) (RR: 0.27; 95% CI: 0.25–0.29), 2) those vaccinated with Zostavax (excluding any Shingrix vaccinations) (RR: 0.92; 95% CI: 0.90–0.94), 3) those vaccinated with PCV-13 (excluding any PPSV-23 vaccinations) (RR: 0.73; 95% CI: 0.71–0.74), and 4) those vaccinated with PPSV-23 (excluding any PCV-13 vaccinations) (RR: 0.71; 95% CI: 0.69–0.73) when compared to unvaccinated groups. The distributions of follow-up time (from start of follow-up to AD onset, death, or censoring) for each of the analyses are shown in Supplementary Table 3. For the vaccinated groups, the follow-up time began when the first target vaccine was received during the follow-up period.

Sensitivity analysis: Controlling for healthy adherer bias

After excluding patients with missing demographics, 1,530,385 patients were tentatively eligible for the sensitivity analysis; among this group, 544,228 had at least one statin medication record in the first half of the look-back period (i.e., 2009-2010). Of those patients, 281,554 had a PDC ≥ 80% during the second half of the look-back period (2010–2011) and were therefore eligible for matching and ATT estimation. Similar to the primary analyses, analyses of this subset of statin adherers revealed statistically significant reductions in AD risk after vaccination: Tdap/Td vaccination (RR: 0.67; 95% CI: 0.64–0.71), HZ vaccination (RR: 0.71; 95% CI: 0.68–0.73), and pneumococcal vaccination (RR: 0.73; 95% CI: 0.70–0.75). A comparison between the sensitivity analysis results and the main results are displayed in Table 4.

Mechanisms and vaccine types

The mechanisms that underlie the reduced incidence of AD through vaccinations in our cohort need to be explored further. There may be mitigation of disease-specific mechanisms through the prevention of the disease (e.g., herpes zoster) or the reduction in the severity of the disease that have a diminishing effect on the risk of AD. However, because the results from our previous study with influenza vaccination [38] and now the results from this study demonstrate that multiple vaccinations are associated with a reduced incidence of AD, it may be that there are other, more general mechanisms. These other mechanisms could include innate immune system training and lymphocyte-mediated cross-reactivity, descriptions of which are both expanded upon in our previous influenza vaccination manuscript [38].

Another factor that should be considered is the age of patients when they receive their vaccines against tetanus and diphtheria, with and without pertussis; herpes zoster; and, pneumococcus. The immunogenicity of vaccines is reduced in patients as they age, therefore there is a decrease in vaccine efficacy [49]. Analyses in Supplementary Figure 1A-C illustrate that the incidence of AD increases with age; however, the risk of developing AD is still diminished in association with the use of Tdap/Td (Supplementary Figure 1A), HZ (Supplementary Figure 1B), and pneumococcal (Supplementary Figure 1 C) vaccinations. As a result, it appears to be advantageous for people 65 years and older to receive these vaccinations to prevent disease and to reduce the risk of AD. Vaccines have been created and have been shown to provide a more robust immune response in patients 65 years and older, including recombinant and conjugated vaccines.

Herpes Zoster: Live-attenuated versus recombinant

Two HZ vaccines have been approved for use in the United States. Zostavax was recommended from 2008–2020. Like the vaccines against varicella recommended in children for protection against primary varicella infection, Zostavax contains a live-attenuated form, but at a much higher titer than currently approved pediatric varicella vaccines [50]. Shingrix, on the other hand, is a recombinant vaccine against HZ that contains both the varicella-zoster glycoprotein E (gE) antigen and the AS01_B adjuvant system [51]. The vaccine utilizes gE as an antigen since it is the glycoprotein that varicella-zoster exhibits most frequently; this glycoprotein is also the target for varicella-zoster CD4+ T cell response [51]. Both Zostavax and Shingrix are capable of eliciting T-cell-independent and T-cell-dependent responses; however, the efficacy of protection provided by these two vaccines differs significantly. The efficacy of Zostavax in HZ risk reduction was only slightly over 50% in patients 60 years and over with previous varicella zoster infection, and the HZ protection provided by this live vaccine reduced after approximately five years [52]. An advantage to Zostavax was that it was given as a one-time dose. Shingrix, in contrast, has an efficacy of over 90% in reducing HZ risk and, unlike Zostavax, can be safely administered to immunocompromised patients [15, 52]. Shingrix is administered over two doses, with protection lasting approximately seven years [21]. It is now recommended by the CDC that those who previously received Zostavax also receive Shingrix [7].

Pneumococcal: Polysaccharide versus conjugated

For the unconjugated polysaccharide vaccine (i.e., PPSV), the antigenic component consists of polysaccharides from the capsule of pneumococcus [25]. These vaccines can only produce a limited immune response because the polysaccharides are unable to be loaded into the major histocompatibility complex (MHC) cavity; therefore, although they elicit production of IgM antibodies by B cells, polysaccharide vaccines cannot induce T-cell-dependent responses and thus lack several effects of peptide-containing vaccines, including the production of memory B cells, antibody class switching, or affinity maturation [37]. In contrast, conjugated vaccines incorporate capsular polysaccharides covalently bound to a carrier protein in order to elicit a more robust immune response [25]. For PCV13, the carrier is a genetically detoxified form of the diphtheria toxin protein [53]. The conjugate allows both the polysaccharide and the carrier protein to be loaded into the MHC-II cavity, thus allowing for activation of helper T cells [37]. This T-cell-dependent pathway enables the production of memory B cells and non-IgM antibodies (e.g., IgG, IgE). Therefore, the PCV is thought to have a more sustained immune response, overall, when compared with PPSV. The current recommendations have expanded the use of PCV vaccinations. PCV15 and PCV20 were approved by the FDA in 2021. It is now recommended that patients 65 years and older receive either a dose of PCV20, or a dose of PCV15 followed by a dose of PPSV23 one year later.

Public health and an addition to a clinician’s toolkit

This study suggests that it is important for patients to have ready access to routine adult vaccinations. Over the past 15 years there has been an incremental increase in vaccine coverage every year for vaccines preventing tetanus or diphtheria, with and without pertussis; herpes zoster; and pneumococcus among adults in the United States [54]. For example, from 2008 to 2018, the rate of patients who received an HZ vaccine increased significantly from 6.7% to 34.8% [55]. Also, it is estimated that 58.9% of adults 65 and older were exposed to a tetanus-containing vaccine between 2008 and 2018 [54]. The increase continued until the COVID-19 pandemic and subsequent shutdowns. During this period, there were reductions in the administration of adult vaccines, with the HZ vaccination rates dropping by 89% and Tdap/Td rates by 70% [56]. Despite the shutdowns and physical isolation, elderly patients are still at risk for developing HZ because the disease is caused by a reactivation of varicella-zoster, as opposed to a new microbial exposure [57]. This reactivation is also associated with an increase in dementia risk [18]. It is estimated that 3.9 million HZ vaccinations were missed in 2020 due to COVID-19 shutdowns, accounting for an estimated 31,945 additional HZ cases over two years [57]. It is not yet known whether the decrease in vaccination coverage and an increase in vaccine preventable diseases will affect dementia rates.

The clinician-patient relationship, as well as the understanding and knowledge of vaccinations are important parts of a patient’s decision to refuse or accept a vaccine [58]. The value of vaccination, as we have demonstrated, goes beyond preventing infection or severe disease from that infection. In fact, there are multiple non-specific potential benefits of vaccination such as improving asthma severity [59], AD prevention [38, 48], and use as an adjuvant cancer therapy (even though it is administered through a non-traditional route) [60, 61], among others. Nicholls et al. [62] found that by emphasizing disease susceptibility and vaccine efficacy/benefits, patients may be more willing to receive vaccinations in the future. By discussing these added non-specific advantages of vaccination with patients, clinicians may be able to convince hesitant patients that the benefits of vaccination with one of the routine adult vaccinations outweighs the risks.

Sensitivity analysis

In order to assess the extent to which healthy adherer bias influenced our results, we performed a similar sensitivity analysis to Wiemkem et al. [30] in which we only included patients who were adherent to statin medications. Because the results from the sensitivity analysis were similar to those results within the original main analysis, we concluded that our study findings showing the association between exposure to adulthood vaccinations and a decreased incidence of AD were not influenced by healthy adherer bias.

Limitations

There are several limitations to our study. 1) Optum’s CDM only includes patients with both medical and prescription coverage. Therefore, those with medical insurance but no prescription coverage and vice versa were not included in this study, limiting the generalizability of our findings. The CDM may also lack vaccine exposures for patients who pay out of pocket for their vaccinations; however, if patients were to use their insurance card for vaccinations, then their vaccination would be recorded. 2) Because our study is retrospective in nature, and the main objective for data collection was not adult vaccinations and AD diagnosis, there is risk for misclassification bias. 3) For the outcome variables, we attempted to control for misclassification by including patients that had no AD-related diagnoses or medications or that had at least two healthcare records with some combination of AD-related diagnoses or medications; patients with only one AD-related diagnosis or medication record were excluded to minimize misclassification due to clerical errors. Furthermore, we included the diagnosis codes for “senile” or unspecified dementia in our case-identification algorithm for AD because, although the majority of clinical dementia among older adults is secondary to AD, the prevalence of AD-specific diagnostic codes in Medicare claims data is far lower (and the rate of unspecified dementia diagnostic codes far higher) than would be expected based on the known preponderance of AD as the underlying cause of dementia in this population [41]. We do not, however, know the true rates of with dementia secondary to AD versus other causes of neurodegeneration in the CDM. 4) Another consideration and potential limitation of this study was the decision to count vaccinations as valid exposures as long as they occurred at least one day before the initial AD diagnosis. 5) The risk of immortal time bias is another important consideration in this study. To provide a measurement of the time at-risk among vaccinated patients that does not include the period of “immortality” they experience between the start of the follow-up period and the date of vaccination, the distribution of follow-up duration (Supplementary Table 3) for vaccinated patients was defined as the time from vaccine receipt (rather than the start of the follow-up period) to date of incident AD, death, or censoring (i.e., the patient’s last record before the end of the follow-up period). As shown in Supplementary Table 3, the median at-risk period for the vaccinated group was greater than that of the unvaccinated group in most of the analyses, a disparity that should be considered when interpreting the results of this study. 6) Although the SMD for each of the post-PSM covariates was ≤ 0.10, which meets the conventional definition for adequate covariate balance between the vaccinated and unvaccinated groups [44], the presence of higher disease burden within the vaccinated groups is noted. If there is a bias present from this difference in comorbidity distributions, it would predispose our analysis against finding a protective effect. 7) While our study did control for some sociodemographic and comorbid conditions, we could not control for other behaviors and characteristics that may influence vaccination acceptance or refusal, such as marital status, educational level, and income status [58, 62]. We reported E-values for each of the point estimates to provide an estimate of how strongly an unmeasured confounder would need to be associated with both the exposure and outcome (adjusting for the same covariates as this analysis) in order to render the point estimate statistically insignificant. 8) Moreover, some vaccines were approved and recommended for use in the general population during our study period. Shingrix is an example: it was introduced in 2017, two years before the end of our study period. While we were able to move the follow-up period to start in 2017, this did result in a limited period of follow-up (2 years) for patients to receive Shingrix and to study its impact on AD incidence. 9) Finally, exposure to diseases such as HZ and influenza have been associated with an increased incidence of AD; however, we did not control for this in our models because of the difficulty in obtaining an accurate diagnosis for infections, such as influenza, which may lead to misclassification. Relatedly, we cannot be certain whether our observations relate to reduced infection rates versus vaccine-related effects on the immune system.