The heterogeneous treatment effects of statins on dementia: a target trial emulation with causal machine learning using integrated genetic and real‐world data

Abstract

INTRODUCTION

Given the complexity of dementia, the inconsistent evidence on statins and dementia highlights the need for robust methods to assess heterogeneous treatment effects (HTEs).

METHODS

We emulated a target trial using UK Biobank comparing statin initiators and non‐initiators aged ≥55 years. Marginal structural models were fitted to estimate 5‐year adjusted risk difference (aRD). We used iterative causal forest, a causal machine learning subgrouping algorithm, to identify subgroups with HTEs.

RESULTS

Among 18,366 participants, the overall aRD for all‐cause dementia was −1.0‰ (95% CI: −4.2‰ to 2.3‰). We identified subgroups by polygenic risk score for Alzheimer's disease (AD) excluding apolipoprotein E (APOE) genotype (“non‐APOE PRS”). Participants with high non‐APOE PRS showed cognitive benefit (all‐cause dementia: aRD −5.9‰, 95% CI: −8.1‰ to 1.2‰; AD: aRD −5.0‰, 95% CI: −8.2‰ to −0.2‰).

DISCUSSION

Participants with high non‐APOE PRS may benefit from statins, suggesting genetic susceptibility beyond APOE could modify statins' cognitive effects.

Highlights

• Sustained statin use did not reduce the 5‐year risk of dementia in the overall population.

• Heterogeneous treatment effect analysis indicated a potential cognitive benefit among participants with high PRSs for dementia excluding APOE genotype.

• Statins may benefit genetically susceptible individuals independent of APOE genotype, supporting precision prevention strategies.

1. BACKGROUND

Dementia is a progressive neurodegenerative disorder characterized by cognitive deterioration and functional impairment, with Alzheimer's disease (AD) being the most prevalent underlying pathology. It affects over 55 million people globally, significantly burdening healthcare systems and caregivers. ¹ Despite its impact, treatment options remain limited. Four symptom‐relief drugs (donepezil, rivastigmine, galantamine, and memantine) and two disease‐modifying agents (anti‐amyloid monoclonal antibodies: lecanemab and donanemab) are available. ² Although monoclonal antibodies have shown promise in reducing amyloid beta (Aβ) and slowing cognitive decline in early AD, their long‐term efficacy and safety remain uncertain, ³ underscoring the urgent need for novel treatment and drug repurposing (identifying new indications of on‐market drugs). ⁴

Elevated low‐density lipoprotein cholesterol (LDL‐C) during midlife constitutes an established modifiable risk factor for dementia, ⁵ prompting interest in statins – potent inhibitors of cholesterol biosynthesis – as potential therapies. Nevertheless, their neurological effects remain debated. Two large‐scale trials found no clear cognitive benefits, ⁶ , ⁷ although these studies had significant limitations, such as secondary cognitive endpoints and lack of generalizability (inclusion of individuals with existing vascular disease ⁶ or high 5‐year risk of death from coronary heart disease ⁷ ). Previous cohort studies ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ , ¹⁴ yielded inconsistent findings (Table S1). These studies are limited by critical methodological flaws such as inadequate power for dementia outcome, ⁸ , ⁹ unclear exposure definitions, ¹¹ prevalent user bias, ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² and asymmetrical follow‐up designs (allowing switching from non‐statins to statins group but not vice versa). ¹⁰ , ¹⁵

Moreover, previous studies focused on the average treatment effect in the treated (ATT) of statins in the overall population. ⁹ , ¹² , ¹⁶ Given dementia's complex biological pathways and older adults’ heterogeneity, the effect of statins likely varies across elderly subpopulations with varying genetic factors, comorbidities, and socioeconomic conditions. ¹⁷ Thus, evaluating heterogeneous treatment effects (HTEs) to identify subgroups with the greatest cognitive benefits is essential. ¹³ Previous studies ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² inadequately addressed HTEs, primarily relying on traditional one‐variable‐at‐a‐time subgroup analysis (with the exception of Li et al., ⁸ who evaluated a statin–age–apolipoprotein E [APOE] genotype interaction) (Table S1). These traditional approaches are limited by small sample size, confounding, and dependence on prior knowledge. ¹⁸ , ¹⁹ Recent advances in causal machine learning‐based subgrouping algorithms, ²⁰ , ²¹ notably the iterative causal forest (iCF), overcome these limitations, enabling subgroup identification in observational data while controlling confounding without requiring prior knowledge. ²² , ²³

Genetic factors influence both dementia risk and variability in drug response. ¹⁷ The APOE ε4 allele – the strongest genetic risk factor for dementia – has been associated with enhanced cognitive benefit from statins in some studies, ²⁴ , ²⁵ , ²⁶ with a reported AD hazard ratio (HR) of 0.60 for APOE ε4 carriers compared to 0.96 for non‐carriers during a median follow‐up of 9.8 years (p for interaction = 0.015). ²⁵ However, the role of dementia‐related genetic variants beyond APOE region remains understudied. ²⁷ Genes influencing LDL‐C response to statins, such as those involved in statin metabolism pathways (e.g., cytochrome P450), may also impact cognitive outcomes. ¹⁷ , ²⁸ , ²⁹ To guide personalized therapy strategies, rigorous designs combined with advanced HTE analysis are necessary to evaluate both genetically predicted dementia risk and statin‐induced LDL response.

According to the US Food and Drug Administration, real‐world data (RWD) are “data relating to patient health status and/or the delivery of healthcare routinely collected from a variety of sources.” ³⁰ The increasing availability of RWD‐linked biobanks offers opportunities to integrate genetic data with RWD, enhancing our understanding of genetic influences on drug effectiveness. ³¹ Leveraging UK biobank (UKB) genetic data and RWD, this study employs causal inference methodologies – including target trial emulation and causal machine learning subgroup analysis – to identify the subpopulation that benefits most from statins on dementia outcomes. We hypothesize that genetic variation contributes to heterogeneity in statins’ effect on dementia, aiming to inform personalized and more effective treatment strategies.

2. METHODS

2.1. Data source

The data source was UKB, a prospective cohort study that encompasses over 500,000 participants from the UK. Detailed information about UKB has been described elsewhere. ³² To facilitate longitudinal analysis, UKB is dynamically linked to various electronic health administrative datasets. These datasets, representing a rich source of RWD, include hospitalization records, death registry, and primary care encounters. As of January 1, 2025 (date of data analysis initiation), primary care data concluding at 2016–2017 (May 31, 2016, for England [TPP supplier], May 25, 2017, for England ^{[Vision supplier]}, March 31, 2017, for Scotland, and August 31, 2017, for Wales) was accessible for a subset comprising 45% of UKB participants. ³³ The data included diagnoses, laboratory tests, and prescriptions. ³⁴ To ensure methodological rigor in pharmacotherapeutic data processing, medications obtained from the prescription records were mapped to Anatomical Therapeutic Chemical classification system via manual coding by a 12‐member clinical pharmacist team, led by author Y. Z. and supervised by author Y.X., which harmonized multisource identifiers (drug name, Read version 2 code, British National Formulary code, Dictionary of Medicines and Devices code) through dual independent verification with discrepancies resolved by discussion.

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using PubMed. Previous real‐world studies about the cognitive effects of statins on dementia have yielded inconsistent results. Moreover, there is limited understanding of which specific patient subgroups may benefit most from statin use. The role of genetic factors in modifying the cognitive effects of statins remains largely understudied.

2. Interpretation: Our study found that sustained statin use did not reduce the risk of dementia in the overall population. However, among individuals with a high PRS for dementia that excludes the APOE genotype (“non‐APOE PRS”), statin use was associated with a potential reduction in dementia risk.

3. Future directions: These findings may help identify subpopulations most likely to benefit from statins for dementia prevention, supporting the development of precision medicine strategies.

2.2. Target trial emulation and heterogenous treatment effect quantification

Utilizing the rich RWD and genetic data from UKB, we emulated a target trial to (1) estimate the ATT of statins on the risk for dementia and (2) quantify the HTEs using the iCF algorithm. A graphical overview of the study design is presented in Figure 1 and Figure 2, and the key elements of the target trial protocol and its emulation process are detailed below and in Table S2.

FIGURE 1.

Processing pipeline for real‐world and genetic data from UK Biobank. ATC, Anatomical Therapeutic Chemical classification system; PRS, polygenic risk score; SNP, single‐nucleotide polymorphism; UKB, UK Biobank.

FIGURE 2.

Overview of the target trial emulation study with iCF analysis to assess the ATT and HTEs of statins on dementia. ATT, average treatment effect in the treated; APOE, apolipoprotein E; HTE, heterogeneous treatment effect; IPCW, inverse probability of censoring weight; iCF, iterative causal forest; LDL‐C, low‐density lipoprotein cholesterol; MSM, marginal structural model; PRS, polygenic risk score; PS, propensity score; RWD, real‐world data; UKB, UK Biobank.

2.2.1. Study population and design

Participants included were aged 55 years and older with linked primary care data. Exclusion criteria included non‐White ethnicity, statin or non‐statin cholesterol‐lowering medication prescriptions in the prior 2 years, a history of mild cognitive impairment (Table S3), dementia diagnosis or treatment (donepezil, rivastigmine, galantamine, and memantine, detailed in Table S4), diagnosis of statin contraindication (i.e., active liver disease), ³⁵ with ambiguous APOE genotype, missing imputed genome‐wide genotype data or baseline LDL‐C measurements.

UKB recruited participants between 2006 and 2010, and over half of participants were enrolled after January 2009. We used a sequential trial design by setting the index date (T₀) as the first day of each month from January 2009 to December 2011. Each corresponding month served as the enrollment period, generating 36 sequential trials to expand the pool of eligible participants. The same inclusion and exclusion criteria were consistently applied across 36 trials, including no prior statin use, allowing participants meeting the criteria to be enrolled in multiple trials. An overview of the sequential design is presented in Method S1.

2.2.2. Treatment strategies

In statin clinical trials, treatment status may change due to clinical reasons such as the onset of a serious illness, prompting clinicians to initiate or discontinue statin based on competing clinical priorities. Thus, we followed a previously used approach to assess the effect of sustained statin use versus no use, defining two treatment strategies: “initiate statin treatment at baseline and remain on it during the follow‐up unless a serious illness (cancer or heart disease) occurs” versus “refrain from statin treatment during the follow‐up unless a serious illness occurs.” ³⁶

2.2.3. Treatment assignment

We classified participants into one of two treatment strategy groups according to the mapped prescriptions (Table S5) within the enrollment period. Statin initiators included those prescribed atorvastatin, rosuvastatin, simvastatin, pravastatin, or fluvastatin; cerivastatin was excluded due to its market withdrawal. Non‐initiators remained eligible for inclusion in all consecutive monthly trials until they initiated statin therapy.

2.2.4. Outcomes

The primary outcome was all‐cause dementia and secondary outcome was AD. To improve sensitivity, dementia cases were identified using the earliest occurrence from two data sources. The first source was primary care data, where Read codes (Table S6) compiled by Wilkinson et al. ³⁷ were used to identify dementia, with reported positive predictive values of 86.8% for all‐cause dementia and 74.1% for AD. The second source was algorithmically defined health‐related outcomes developed by UKB outcome adjudication group (data fields 42020 and 42018), which integrate self‐reported diagnoses, hospital admissions, and death registry data to detect dementia. ³⁸

2.2.5. Follow‐up

Participants were followed from the last day of the enrollment period of each emulated trial until the first occurrence of study outcomes, death, loss to follow‐up (defined by data field 191), or the end of primary care data.

2.2.6. Causal contrast

The intention‐to‐treat (ITT) effect reflects the effect of treatment initiation irrespective of subsequent adherence, whereas the per‐protocol (PP) effect estimates the effect of sustained use. In our study, many participants deviated from their initial treatment during follow‐up (Table S7), which would tend to bias the ITT effect estimate toward the null. Therefore, our main causal estimand is the PP effect. Participants were further censored at the time of deviation from their assigned treatment: statin initiators were censored upon discontinuation (except following cancer or heart disease diagnosis), and non‐initiators were censored upon statin initiation, again with the same exceptions. Treatment discontinuation was defined as the absence of a subsequent statin prescription within 90 days (the grace period) after the estimated end of supply of the previous prescription. The supply of each statin prescription was estimated using the dose per tablet, the number of tablets, and the defined daily dose, ³⁹ which reflects the assumed average maintenance dose per day for the drug's main indication in adults (Method S2 and Table S8).

2.2.7. Confounders

Through a comprehensive literature review, ³⁶ , ⁴⁰ , ⁴¹ , ⁴² , ⁴³ , ⁴⁴ two sets of confounders were extracted for further adjustment: baseline and time‐varying confounders. Baseline confounders, extracted at T₀ and kept constant during the follow‐up, included age, gender, primary care data provider, APOE genotype, Townsend deprivation index, education level, physical activity, alcohol consumption, smoking status, dietary habit, social isolation, residential air pollution, and reaction time (to reflect baseline cognitive function). Time‐varying confounders, extracted at T₀ and updated at each month during the follow‐up, included body mass index (BMI), high‐density lipoprotein cholesterol, LDL‐C, triglycerides, comorbidities (e.g., stroke, heart disease, and cancer), medication history (e.g., antihypertensives, anticholinergic drugs, ⁴² and menopausal hormone therapy drug ⁴⁴ ). Detailed definitions of the confounders are provided in Tables S9‐S18.

2.2.8. Effect modifiers

Two broad categories of potential effect modifiers were extracted at T₀ for each emulated trial (Figure 2 and Table S9): environmental modifiers and genetic modifiers. Environmental modifiers encompassed sex, age, education level, physical activity, alcohol intake, smoking status, dietary habit, social isolation, residential air pollution, baseline reaction time, baseline BMI, baseline LDL‐C levels, and all previously listed baseline comorbidities; while genetic modifiers included genotype and polygenic risk scores (PRSs) reflecting dementia susceptibility and the LDL‐C‐lowering response to statin treatment, as detailed in Method S3 and Tables S19 and S20. Dementia susceptibility was assessed using both the APOE genotype (grouped as APOE ε4 carriers or non‐carriers) and a PRS for AD excluding the APOE region (“non‐APOE PRS”). The non‐APOE PRS, developed by Ebenau et al., ²⁷ consisted of 39 single‐nucleotide polymorphisms (SNPs) predicting AD independently of APOE genotype. The statin response was assessed using a 35‐SNP PRS (“statin PRS”) established by Mayerhofer et al. ⁴⁵ They reported that a higher statin PRS (per 1 standard deviation [SD] increase) correlated with a greater reduction in LDL‐C levels (−0.05 mg/dL/year) after statin treatment. All the PRSs were then divided into tertiles, and participants were categorized into two genetic risk groups: low (PRS ≤ second tertile) and high (PRS > second tertile).

2.2.9. Estimation of ATT

For each emulated trial, separate 1:1 propensity score (PS) matching without replacement was conducted based on all the baseline confounders and time‐varying confounders extracted at T₀ using the nearest‐neighbor method with a caliper of 0.2 SD of the logit of the PS. Confounder balance was assessed by absolute standardized mean differences (ASMDs), with values exceeding 0.1 indicating meaningful imbalance. To increase the number of outcome events, PS‐matched cohorts from 36 emulated trials were stacked together into an analytical cohort (Figure 2). Furthermore, inverse probability of censoring weighting (IPCW) was then applied to account for the selection bias induced by differential censorship between treatment arms (informative censoring), ³⁶ which was detailed in Method S4.

On the analytical cohort, we fitted a marginal structural model (MSM) to estimate the PP effects of following the assigned treatment strategies on the study outcome. MSM was fitted by pooled logistic regression with treatment strategy, month, month squared, and weighted by IPCW. ⁴⁶ Weights were truncated at the 0.05th and 99.95th percentiles before fitting the MSMs. Adjusted HRs were directly derived from the MSMs. Predicted probabilities from the pooled logistic model were used to estimate adjusted 5‐year absolute risks of study outcome under each treatment strategy and the corresponding 5‐year adjusted risk difference (aRD). Furthermore, weighted cumulative incidence curves were constructed to visualize temporal comparisons of adjusted absolute risks of study outcome between treatment strategies. For all weighted analyses, 95% confidence intervals (CIs) were estimated using non‐parametric bootstrapping with 500 samples.

2.2.10. HTE estimation

The iCF algorithm ²³ was employed on the analytical cohort to identify subgroups with HTEs (Figure 2 and Method S5). Building on the causal forest framework, this data‐driven method iteratively grows forests of varying depths using selected important effect modifiers (with variable importance value greater than the mean), selects stable and explainable decision tree structures (as subgroup decisions) through a plurality voting mechanism at different depths, and then obtains cross‐validated subgroup decisions that best predict treatment effect. Unlike traditional subgroup analysis that relies on predefined models, iCF enables the identification of complex, multivariable interactions without prior assumptions. The non‐parametric, voting‐based features of iCF enhance stability and minimize biases from model misspecification.

In this analysis, the algorithm parameters were set to 1000 trees and 100 iterations for subgroup identification. For each identified subgroup, conditional average treatment effects (CATEs) were quantified on a risk difference scale, as absolute risk is preferred measures for subgroup‐specific treatment effect and is also the scale used by the causal forest. ⁴⁷ , ⁴⁸ Within each identified subgroup, adjusted 5‐year absolute risks of study outcome under each treatment strategy and the corresponding 5‐year aRD were estimated via inverse probability of treatment weight (IPTW)‐IPCW weighted MSM (Method S4). Additionally, when residual confounding was present (indicated by largest weighted ASMD > 0.1), we applied asymmetric PS trimming to exclude participants treated most contrary to prediction, using a cut point at the 0.05th and 99.95th percentiles of the PS distribution for the treated and untreated participants, respectively, increasing thresholds if necessary. ²² , ⁴⁹ Within each identified subgroup by iCF, we applied the same PP analysis for CATE as in our primary analysis.

In addition, to demonstrate HTEs across continuous non‐APOE PRS and statin PRS distributions, we graphically represented the 5‐year aRD as a function of PRS by adding a restricted cubic spline of PRS (three knots at the 10th, 50th, and 90th percentiles) and an interaction term of treatment and the restricted cubic spline of PRS into the previous MSM. Meanwhile, we also performed traditional subgroup analyss to assess for effect modification by key environmental modifiers. These included age and sex, as well as other variables identified as having high variable importance in the raw causal forest (e.g., heart disease, alcohol consumption, physical activity).

2.2.11. Sensitivity analysis

To assess the robustness of our findings, we conducted several sensitivity analyses. First, acknowledging that 1:1 PS matching could reduce the precision of effect estimates, ⁵⁰ we repeated the analyses with a 1:2 PS matching ratio. Second, recognizing that ITT analysis may reduce selection bias during follow‐up and increase the probability of achieving a complete 5‐year follow‐up (aligning well with the causal forest's risk difference scale), we conducted additional analyses on an ITT analytical cohort followed for a maximum of 5 years. To further ensure a 5‐year follow‐up window, we required the ITT cohort entry dates to be at least 5 years before administrative censoring. Third, to explore the effect of statin type, we restricted the sample to simvastatin initiators, as this subgroup comprised over 80% of the treated cohort, and then repeated the analyses. Fourth, to more comprehensively investigate potential effect modification, we incorporated two additional classes of potential modifiers into the iCF algorithm: (1) baseline concomitant medications and (2) predefined SNPs in statin pharmacokinetic and pharmacodynamic pathways (Table S21). Fifth, we compared the iCF analysis results with those from two alternative subgrouping methods: (1) aVirtualTwins method and (2) logistic regression with a Least Absolute Shrinkage and Selection Operator (LASSO) penalty for interaction discovery (Method S6).

PLINK version 2.0 was used for SNP extraction and PRS calculation; all other analyses were performed using R version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria).

3. RESULTS

3.1. Participant characteristics

Following application of the inclusion and exclusion criteria, 9193 statin initiators and 2,339,773 non‐initiators from 36 emulated trials were included (Figure S1). Characteristics of these participants both before and after PS matching are summarized in Table 1. Among statin initiators, 7424 participants (80.8%) were prescribed simvastatin, followed by atorvastatin (1205 [13.1%]), pravastatin (379 [4.1%]), rosuvastatin (160 [1.7%]), and fluvastatin (15 [0.2%]). Before matching, compared to non‐initiators, statin initiators were older (mean age 63.5 vs 62.3 years), had a higher proportion of males (58.4% vs 41.2%), and showed elevated baseline LDL‐C levels (mean 4.08 vs 3.72 mg/dL). APOE ε4 carrier status did not differ between the two groups (25.3% vs 27.8%, ASMD = 0.058). The prevalence of hypertension (46.9% vs 28.8%), diabetes (6.0% vs 2.3%), and heart disease (6.4% vs 2.7%) was also higher in statin initiators, as was the use of other cardiovascular medications. After PS matching, a total of 18,366 participants (9183 statin initiators and 9183 non‐initiators) with well‐balanced characteristics were included in subsequent analyses (Table 1, Figure S2). The PS distributions demonstrated a sufficient overlap between statin initiators and non‐initiators (Figure S3).

TABLE 1.

Baseline characteristics of participants from 36 emulated trials before and after PS matching.

	Unmatched			PS matched
	Non‐initiator (N = 2,339,773)	Statin initiator (N = 9193)	ASMD	Non‐initiator (N = 9183)	Statin initiator (N = 9183)	ASMD
Demographic characteristic
Male, n (%)	964,051 (41.2)	5369 (58.4)	0.349	5464 (59.5)	5361 (58.4)	0.023
Age, years, mean (SD)	62.27 (4.54)	63.46 (4.38)	0.266	63.08 (4.58)	63.46 (4.38)	0.084
Primary care data provider, n (%)			0.285			0.024
England (vision)	198,405 (8.5)	867 (9.4)		844 (9.2)	867 (9.4)
Scotland	221,518 (9.5)	250 (2.7)		219 (2.4)	250 (2.7)
England (TPP)	1,733,203 (74.1)	7282 (79.2)		7307 (79.6)	7274 (79.2)
Wales	186,647 (8.0)	794 (8.6)		813 (8.9)	792 (8.6)
Townsend deprivation index, n (%)			0.105			0.025
Least deprived	457,891 (19.7)	1539 (17.0)		1647 (17.9)	1565 (17.0)
Intermediate deprivation	1,384,782 (59.7)	5275 (58.4)		5279 (57.5)	5372 (58.5)
Most deprived	478,379 (20.6)	2214 (24.5)		2257 (24.6)	2246 (24.5)
Educational level = high, n (%)	1,069,594 (46.0)	3513 (38.9)	0.145	3497 (38.1)	3580 (39.0)	0.019
Genetic factor, n (%)
APOE ε4 carriers	591,931 (25.3)	2560 (27.8)	0.058	2502 (27.2)	2556 (27.8)	0.013
Lifestyle factor, n (%)
Physical activity = enough	1,164,578 (51.1)	4329 (49.2)	0.038	4585 (49.9)	4501 (49.0)	0.018
Alcohol consumption			0.094			0.030
Abstinence	366,380 (15.8)	1472 (16.3)		1432 (15.6)	1495 (16.3)
1 to 14 units/week	1,200,523 (51.7)	4265 (47.2)		4265 (46.4)	4330 (47.2)
>14 units/week	755,396 (32.5)	3296 (36.5)		3486 (38.0)	3358 (36.6)
Smoking status			0.205			0.086
Never	1,253,684 (54.2)	4071 (45.2)		4119 (44.9)	4164 (45.3)
Previous	875,066 (37.8)	3796 (42.1)		4133 (45.0)	3861 (42.0)
Current	186,107 (8.0)	1141 (12.7)		931 (10.1)	1158 (12.6)
Dietary habit = healthy	1,208,504 (52.0)	4219 (46.7)	0.107	4264 (46.4)	4288 (46.7)	0.005
Social isolation = yes	436,143 (18.9)	1734 (19.3)	0.011	1764 (19.2)	1776 (19.3)	0.003
Residential air pollution = high	699,916 (33.4)	3011 (34.0)	0.013	3122 (34.0)	3107 (33.8)	0.003
Comorbidity, n (%)
Stroke	23,488 (1.0)	210 (2.3)	0.101	178 (1.9)	207 (2.3)	0.022
Hypertension	674,981 (28.8)	4311 (46.9)	0.379	4308 (46.9)	4305 (46.9)	0.001
T2DM	54,269 (2.3)	553 (6.0)	0.186	478 (5.2)	548 (6.0)	0.033
Heart disease	62,831 (2.7)	586 (6.4)	0.178	542 (5.9)	582 (6.3)	0.018
PD	5584 (0.2)	27 (0.3)	0.011	20 (0.2)	27 (0.3)	0.015
Depression	233,739 (10.0)	1003 (10.9)	0.030	1013 (11.0)	999 (10.9)	0.005
CKD	62,412 (2.7)	534 (5.8)	0.156	480 (5.2)	531 (5.8)	0.024
Liver disease	17,707 (0.8)	95 (1.0)	0.029	90 (1.0)	95 (1.0)	0.005
Hearing loss	611,674 (26.1)	2614 (28.4)	0.051	2645 (28.8)	2611 (28.4)	0.008
Traumatic brain injury	5512 (0.2)	28 (0.3)	0.013	26 (0.3)	26 (0.3)	<0.001
Cancer	293,195 (12.5)	1138 (12.4)	0.005	1141 (12.4)	1136 (12.4)	0.002
Medication history, n (%)
Antihypertensives	32,474 (1.4)	245 (2.7)	0.091	219 (2.4)	244 (2.7)	0.017
Diuretics	169,452 (7.2)	1379 (15.0)	0.249	1381 (15.0)	1377 (15.0)	0.001
Beta blockers	111,304 (4.8)	915 (10.0)	0.200	842 (9.2)	911 (9.9)	0.026
Calcium channel blockers	151,508 (6.5)	1376 (15.0)	0.277	1357 (14.8)	1375 (15.0)	0.006
RAAS inhibitors	234,301 (10.0)	2151 (23.4)	0.365	2114 (23.0)	2146 (23.4)	0.008
Anticholinergic drug	212,504 (9.1)	1100 (12.0)	0.094	1067 (11.6)	1099 (12.0)	0.011
Menopausal hormone therapy	70,572 (3.0)	895 (9.7)	0.278	787 (8.6)	889 (9.7)	0.039
Aspirin	124,037 (5.3)	331 (3.6)	0.083	309 (3.4)	330 (3.6)	0.012
Other health conditions
Baseline reaction = poor, n (%)	1,406,230 (66.5)	5187 (67.8)	0.029	6185 (67.4)	6236 (67.9)	0.012
BMI, n (%)			0.293			0.022
Not overweight	798,834 (34.5)	2013 (22.1)		2120 (23.1)	2038 (22.2)
Overweight	1,006,039 (43.4)	4311 (47.3)		4272 (46.5)	4340 (47.3)
Obesity	513,217 (22.1)	2782 (30.6)		2,791 (30.4)	2805 (30.5)
HDL‐C, mmol/L, mean (SD)	1.53 (0.40)	1.39 (0.37)	0.374	1.38 (0.35)	1.39 (0.37)	0.003
LDL‐C, mmol/L, mean (SD)	3.72 (0.83)	4.08 (0.96)	0.398	4.08 (0.90)	4.08 (0.95)	0.003
TG, mmol/L, mean (SD)	1.65 (0.90)	2.08 (1.14)	0.411	2.07 (1.14)	2.07 (1.14)	0.004

Abbreviations: APOE, apolipoprotein E; ASMD, absolute standardized mean difference; BMI, body mass index; CKD, chronic kidney disease; HDL‐C, high‐density lipoprotein cholesterol; LDL‐C, low‐density lipoprotein cholesterol; PD, Parkinson's disease; PS, propensity score; RAAS, renin‐angiotensin‐aldosterone system; SD, standard deviation; T2DM, type 2 diabetes mellitus; TG, triglycerides.

3.2. Average treatment effect in the treated

During a median follow‐up time of 5.3 (Q1: 2.2, Q3: 6.6) years, 59 all‐cause dementia events occurred among 9183 statin initiators and 79 among 9183 non‐initiators, corresponding to incidence rates of 1.56 (95% CI: 1.19 to 2.02) and 1.70 (95% CI: 1.34 to 2.11) per 1000 person‐years, respectively (Table 2). The weighted 5‐year absolute risks of all‐cause dementia were 6.8‰ (statin initiators) versus 7.8‰ (non‐initiators), yielding an adjusted risk difference (aRD) of −1.0‰ (95% CI: −4.2‰ to 2.3‰) and an adjusted HR of 0.93 (95% CI: 0.66 to 1.31). For the secondary outcome – AD, the weighted 5‐year absolute risks were 4.9‰ versus 5.8‰, resulting in an aRD of −1.0‰ (95% CI: −3.7‰ to 1.8‰) and an adjusted HR of 0.87 (95% CI: 0.58 to 1.31). The cumulative incidence curves showed an early separation between groups, sustained throughout follow‐up (Figure S4).

TABLE 2.

Number of events, incidence rates, risks, and hazard ratios for study outcomes among statin initiators versus non‐initiators.

Treatment group	No. persons	No. events	Median follow‐up, months (Q1, Q3)	Incidence rate per 1000 person‐years (95% CI)	5‐year absolute risk * , ‰ (95% CI)	5‐year adjusted RD, ‰ (95% CI)	Crude HR (95% CI)	Adjusted HR (95% CI)
Primary outcome: all‐cause dementia
Non‐initiator	9183	79	67.0 (46.3, 79.8)	1.70 (1.34, 2.11)	7.8 (5.6, 10.1)	Reference	Reference	Reference
Statin initiator	9183	59	59.1 (12.8, 76.9)	1.56 (1.19, 2.02)	6.8 (4.5, 9.2)	−1.0 (−4.2, 2.3)	1.55 (1.20, 2.00)	0.93 (0.66, 1.31)
Secondary outcome: AD
Non‐initiator	9183	57	67.0 (46.3, 79.8)	1.22 (0.93, 1.58)	5.8 (3.9, 7.8)	Reference	Reference	Reference
Statin initiator	9183	40	59.1 (12.8, 76.9)	1.06 (0.76, 1.44)	4.9 (2.9, 6.9)	−1.0 (−3.7, 1.8)	1.85 (1.36, 2.52)	0.87 (0.58, 1.31)

Note: ^* Absolute 5‐year risks were estimated using marginal structural models, and this calculation does not imply truncation of follow‐up at the fifth year.

Abbreviations: AD, Alzheimer's disease; CI, confidence interval; HR, hazard ratio; RD, risk difference.

3.3. Heterogeneous treatment effects

For primary outcome – all‐cause dementia, Figure S5 and Table S22 summarized the subgroups identified via iCF by the top 10 important covariates (Figures S6 and S7), which split participants by non‐APOE PRS, APOE genotype, statin PRS, and baseline BMI. Given the small number of events (n < 10) in most iCF‐identified subgroups – which could produce false positives and imprecise CATE estimates ⁵¹ – we combined sparse‐event subgroups split by the same splitting variables to generate larger, more stable subgroups. As shown in Table 3, participants with high non‐APOE PRS (above the second tertile) showed a trend toward benefit from statins use (aRD: −5.9‰, 95% CI: −8.1‰ to 1.2‰), while those with low non‐APOE PRS showed no benefit (aRD: 0.8‰, 95% CI: −1.5‰ to 3.8‰).

TABLE 3.

Adjusted risk differences for all‐cause dementia associated with statin treatment by simplified subgroup decisions from iCF analysis.

Subpopulation identified by iCF algorithm	Cohort	No. persons	No. events	Median follow‐up, months (Q1, Q3)	Incidence rate per 1000 person‐years (95% CI)	Death (%)	5‐year absolute risk (‰)	5‐year adjusted RD (‰) (95% CI)	Largest ASMD after matching or weighting	Median ASMD after matching or weighting
Overall population	Non‐initiator	9183	79	67.0 (46.3, 79.8)	1.70 (1.34, 2.11)	2.8	7.8	−1.0 (−4.2, 2.3)	0.039	0.008
	Statin initiator	9183	59	59.1 (12.8, 76.9)	1.56 (1.19, 2.02)	2.3	6.8
Subpopulation with high non‐APOE PRS	Non‐initiator	3004	36	67.0 (46.3, 79.8)	2.38 (1.66, 3.29)	2.9	12.7	−5.9 (−8.1, 1.2)	0.003	<0.001
	Statin initiator	3101	20	59.1 (12.8, 76.9)	1.58 (0.97, 2.44)	2.4	6.8
Subpopulation with low non‐APOE PRS	Non‐initiator	6179	43	67.0 (46.3, 79.8)	1.37 (0.99, 1.84)	2.8	6.5	0.8 (−1.5, 3.8)	0.001	<0.001
	Statin initiator	6082	39	59.1 (13.8, 76.9)	1.56 (1.11, 2.13)	2.2	7.2

Abbreviations: APOE, apolipoprotein E; ASMD, absolute standardized mean difference; CI, confidence interval; iCF, iterative causal forest; PRS, polygenic risk score; RD, risk difference.

For secondary outcome, AD, the iCF developed by the top 10 important covariates (Figures S8 and S9) split participants by non‐APOE PRS, APOE genotype, and statin PRS (Figure S10 and Table S23). Similar to all‐cause dementia, in the simplified subgroup decision (combining groups with sparse events), participants with high non‐APOE PRS benefited from statins (5‐year aRD: −5.0‰, 95% CI: −8.2‰ to −0.2‰, absolute risk 4.7‰ vs 9.7‰), while those with low non‐APOE PRS showed no benefit (aRD: 0.9‰, 95% CI: −1.0‰ to 3.5‰). Further cross‐validation details of the iCF algorithm are provided in Figures S11–S18 and Tables S24 and S25.

Consistent with subgroup decisions identified by iCF, as non‐APOE PRS increased from −0.0257 to 0.0164, the 5‐year aRD of all‐cause dementia remained near zero below the second tertile threshold (−0.0024) of the non‐APOE PRS distribution but began decreasing noticeably above this point (Figure 3). The relationship between statin PRS and the 5‐year aRD exhibited a reverse J‐shaped curve (Figure 3): Participants with the lowest statin PRS showed a trend toward higher, positive aRD, but this association was not statistically significant; in comparison, less obvious elevated aRD was observed among those with high statin PRS. Similar patterns for non‐APOE PRS, statin PRS, and 5‐year aRD of AD are shown in Figure S19.

FIGURE 3.

PRSs and 5‐year aRD of all‐cause dementia with statin treatment. Note that this figure presents the 5‐year aRD of all‐cause dementia associated with sustained statin treatment. The blue curve represents the estimated aRD, with the shaded blue area indicating the 95% CI. The pink histogram illustrates the distribution of participants across different PRS values. The vertical dashed lines represent the tertiles of the PRS. Panel (A) represents the results for non‐APOE PRS, while panel (B) corresponds to statin PRS. aRD, adjusted risk difference; APOE, apolipoprotein E; CI, confidence interval; PRS, polygenic risk score.

As for environmental modifiers, subgroup analysis revealed that baseline heart disease significantly modified the effect of statins on dementia risk (p for interaction < 0.01), as detailed in Figure 4 and Figure S20. Specifically, participants with baseline heart disease experienced benefits from statin therapy, with a lower risk of all‐cause dementia (aRD: −22.7‰, 95% CI: −42.5‰ to −7.5‰) and AD (aRD: −26.9‰, 95% CI: −42.6‰ to −12.9‰). However, this subgroup represented only about 6% of the analytical cohort (n = 1124) and had few events among statin initiators (n < 10).

FIGURE 4.

Subgroup analyses of association between statins and all‐cause dementia stratified by key environmental factors. CI, confidence interval; HR, hazard ratio; IR/1000PY, incidence rate per 1000 person‐years; RD, risk difference.

3.4. Sensitivity analysis results

For the ATT estimation, the first (1:2 PS‐matched cohort), second (ITT cohort), and third (simvastatin‐restricted cohort) sensitivity analysis (Figures S21–S26) consistently yielded ATT estimates similar to those of the primary analyses, showing no observable difference in the risk of all‐cause dementia or AD with statin initiation in the overall population.

For the HTE estimation, the iCF analyses from the first through fourth sensitivity analyses (the fourth incorporating medication history and additional SNPs as potential effect modifiers; Figures S27 and S28) were generally consistent with the primary analysis. The subgroup partitions were primarily defined by non‐APOE PRS, APOE genotype, statin PRS, age, and alcohol consumption. Specifically, in the first sensitivity analysis based on the 1:2 PS‐matched cohort, the iCF for AD identified subgroups defined by APOE genotype and age (Figure S22). No clear benefit was observed among APOE ε4 carriers (aRD: 0.4‰, 95% CI: −5.0‰ to 5.3‰), whereas APOE ε4 carriers aged ≥65 years showed a modest trend toward lower AD risk (aRD: −6.7‰, 95% CI: −16.8‰ to 6.0‰). The fifth sensitivity analysis, which applied two alternative subgrouping methods (the aVirtualTwins method and logistic regression with a LASSO penalty), identified similar subgroups stratified by non‐APOE PRS or APOE genotype (Table S26 and S27). Some of these subgroups, however, yielded imprecise CATE estimates due to small sample sizes or few outcome events.

4. DISCUSSION

4.1. Statement of principal findings

In this target trial emulation examining statins and dementia risk, no cognitive benefit was observed in the overall population. However, HTE analysis revealed cognitive benefit for dementia among participants with high non‐APOE PRS (all‐cause dementia: aRD: −5.9‰, 95% CI: −8.1‰ to 1.2‰; AD: aRD −5.0‰, 95% CI: −8.2‰ to −0.2‰).

Additionally, participants aged ≥65 years with APOE ε4 allele (AD: aRD: −6.7‰, 95% CI: −16.8‰ to 6.0‰, detailed in the first sensitivity analysis) and those with high statin PRS (all‐cause dementia: aRD: −2.5‰, 95% CI: −8.7‰ to 1.1‰; AD: aRD −2.5‰, 95% CI: −6.3‰ to 2.1‰, detailed in the second sensitivity analysis and Figure 3) or with baseline heart disease (all‐cause dementia: aRD: −22.7‰, 95% CI: −42.5‰ to −7.5‰; AD: aRD −26.9‰, 95% CI: −42.6‰ to −12.9‰, detailed in subgroup analysis) may also benefit from statins.

4.2. Comparison with previous studies

The ATT of statins on dementia has been extensively studied. Observational studies suggest reduced dementia risk with statin use ⁵² , ⁵³ , ⁵⁴ but are limited by selection bias, information bias, and residual confounding. For instance, a UKB study ¹² by Ye et al. reported a higher AD risk among statin users (n = 371,019, 9‐year follow‐up, HR 1.19, 95% CI: 1.08 to 1.30), contrasting with our results. Notably, their reliance on self‐reported medication use at UKB assessment centers introduced prevalent user bias and exposure misclassification. Conversely, we minimized these biases using UKB primary care prescription records. By employing rigorous target trial emulation, our study identified a limited protective effect of statins in the overall population, aligning with two randomized trials ⁶ , ⁷ and three target trial emulation studies. ¹³ , ¹⁴ , ³⁶ The lower 5‐year dementia risks versus Carniglia et al. (non‐initiators: 7.8‰ versus 28‰; statin initiators: 6.8‰ versus 18‰) ³⁶ likely reflect the healthier profile of the UKB cohort. ⁵⁵ Nonetheless, this lack of representativeness is unlikely to substantially bias exposure–disease relationship assessments ⁵⁵ .

Most HTE analyses on statin and dementia focused on APOE genotype, given its prominent role in genetic susceptibility. Several studies suggest stronger protective effects among APOE ε4 carriers. ²⁵ , ⁵⁶ In our study, a similar but non‐significant trend appeared only in participants aged ≥65 years, likely reflecting limited events and follow‐up. Notably, previous studies reporting greater benefits among APOE ε4 carriers were predominantly in populations aged ≥65 years, ²⁵ , ⁵⁶ consistent with our results. Few studies have investigated genetic susceptibility beyond APOE genotype on statin cognitive effects. Notably, Ye et al. ¹² examined the same non‐APOE PRS and suggested harmful effects of statins on AD in individuals with higher scores. However, the aforementioned limitations may bias their estimates.

As for environmental modifiers like sex and age, our findings were consistent with a recent large‐scale HTE study, ¹³ suggesting little to no influence on the statin–dementia association. In contrast, subgroup analysis suggested cognitive benefit among participants with heart disease, consistent with prior studies in heart failure, ⁵⁷ atrial fibrillation, ⁵⁸ and ischemic heart disease. ⁵⁹ This finding should, however, be interpreted with caution given the small subgroup size and limited events.

4.3. Possible explanations and clinical implications

The mechanisms by which statins prevent dementia remain unclear, and the role of genetic modifiers is not fully understood. Dysregulation of cholesterol biosynthesis, transport, and metabolism is increasingly recognized as a contributing factor in AD pathogenesis. ⁶⁰ , ⁶¹ Elevated concentrations of 27‐hydroxycholesterol, a cholesterol metabolite that can cross the blood–brain barrier, have been found to potentially contribute to neurotoxicity and neuronal dysfunction. ⁶² , ⁶³ Statins reduce total cholesterol and are thought to lower 27‐hydroxycholesterol levels by inhibiting cholesterol hydroxylation. ⁶⁴ Additionally, statins exhibit pleiotropic properties beyond cholesterol reduction, including enhancing endothelial function, decreasing oxidative stress and inflammation, attenuating Aβ production and aggregation, stabilizing atherosclerotic plaques, and inhibiting platelet aggregation, all of which may collectively confer neuroprotection and potentially reduce dementia risk. ¹⁷ , ⁶⁵ , ⁶⁶

Our study assessed genetic susceptibility using APOE genotype and non‐APOE PRS. APOE ε4 carriers typically exhibit insufficient levels and impaired function of apolipoprotein E, linked to disrupted cholesterol transport, impaired myelination, and cholesterol accumulation in the brain. ⁶⁷ Similarly, the non‐APOE PRS incorporates SNPs in genes involved in cerebral cholesterol metabolism, such as clusterin (CLU), sortilin‐related receptor 1 (SORL1), and cholesteryl ester transfer protein (CETP). ⁶⁸ Thus, higher non‐APOE PRS may indicate higher genetic risk of brain cholesterol dysregulation, a key feature of dementia pathology. Therefore, the cognitive benefit of statin in individuals with higher genetic susceptibility may be partly explained by the restoration of cholesterol homeostasis in the nervous system. ⁶⁹

This potential mechanism could also explain the protective trend of statins in individuals with higher statin PRS, reflecting stronger genetically predicted LDL‐C‐lowering response. Additionally, higher statin PRS is associated with reduced stroke risk following statin therapy, ⁴⁵ potentially reducing secondary vascular dementia incidence and thus all‐cause dementia risk. ⁶⁹ However, excessive LDL‐C lowering may disrupt cholesterol homeostasis crucial for brain function, potentially causing reversible cognitive impairment. ⁶⁹ , ⁷⁰ Such side effects could sometimes be misdiagnosed as dementia. ⁷¹ This phenomenon may partly explain the “reverse J‐shaped” relationship between statin PRS and all‐cause dementia aRD (Figure 3 and Figure S19). In sensitivity analysis, we also examined six predefined SNPs within statin pharmacokinetic and pharmacodynamic pathways. None emerged in the final subgroup decisions, suggesting their modifying effects may be subtle and less detectable at the population level compared with the combined influence captured by PRS, warranting further investigation.

Our analyses suggest that both genetic and environmental modifiers may influence the cognitive benefit of statins. Although the CATEs in the identified subgroups were not highly precise and the CIs were wide, these subgroups were identified via the rigorous iCF algorithm and trial emulations. The CATEs were derived from a data‐driven, extensive iteration, voting, and cross‐validation process, which enhanced their credibility as potential HTEs. Our findings may help support more targeted dementia‐prevention strategies. For example, individuals with heart disease appear more likely to benefit, suggesting clinicians should consider the presence of underlying heart disease when evaluating the cognitive benefits of statins. The genetically defined subgroup benefits we identified could be used in future clinical practice as genomic data become increasingly integrated into healthcare systems. ⁷² Overall, identifying potential modifiers and incorporating genetic data may enhance our understanding of lipid‐related neurodegenerative pathways, inform the design of future trials targeting those most likely to benefit, and provide a promising avenue for personalized prevention strategies in dementia.

4.4. Strengths and limitations

This study has several strengths. First, we innovatively investigated the influence of genetic susceptibility beyond APOE and genetically predicted drug response on dementia risk. Second, our target trial emulation leveraged the extensive UKB dataset, enabling adjustment for diverse confounders and mitigating biases common in pharmacoepidemiologic studies of dementia, a complex, multifactorial condition. Third, unlike prior UKB‐based studies prone to exposure misclassification, we extracted statin prescription from linked primary care records, precisely identifying new users and treatment discontinuations to ensure a valid new user design and PP analysis. Fourth, our study is the first to apply a robust causal machine learning algorithm within a target trial emulation framework and integrated real‐world genetic data to assess HTEs of statins on dementia; this analytic framework offers an important methodological approach for precision medicine, fully unlocking the potential of RWD‐linked biobanks.

We also acknowledge several limitations. First, our analysis included only participants of European ancestry, and the PRSs used were derived from European populations; thus, findings may not generalize to other populations. Second, the PP analysis required ongoing monitoring of medication exposure, but as prescription data were only available from primary care records, follow‐up had to be truncated at the end of these records (approximately 2016–2017), despite the fact that many dementia diagnoses occurred afterward. This relatively short follow‐up period, combined with the informative censoring inherent in PP analysis, may have reduced dementia events and introduced bias. Nevertheless, this approach better captures real‐world medication patterns compared to prior ITT analyses. ⁷³ Further studies with extended follow‐up (≥10 years) are needed to evaluate the long‐term statin effects. Third, as over 90% of initiators in our cohort used lipophilic statins, results may not generalize fully to hydrophilic statins (e.g., rosuvastatin and pravastatin), warranting further research. Fourth, although we applied rigorous pharmacoepidemiologic methods, the possibility of residual confounding cannot be fully excluded. Fifth, limited events (n < 20) prevented evaluation of vascular dementia. Sixth, dementia diagnoses were based on administrative codes rather than biomarkers or pathology, potentially causing misclassification bias, ⁷⁴ though likely non‐differential between groups. Lastly, we censored death rather than treating it as a competing event, but given the few events, this is unlikely to influence findings.

4.5. Conclusion

Our study found that sustained statin use did not reduce the risk of all‐cause dementia or AD in the overall population. However, HTE analysis indicated cognitive benefit among individuals with high non‐APOE PRS, reflecting dementia risk independent of APOE genotype. Additionally, participants aged ≥65 years with the APOE ε4 allele, with a high statin PRS (indicating stronger genetically predicted LDL‐C lowering), or with baseline heart disease might also benefit from statins. These findings may help identify subpopulations most likely to benefit from statins for dementia prevention, supporting precision medicine strategies.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the supporting information

Supporting information

Supporting Information

Supporting Information

Supporting Information

DATA AVAILABILITY STATEMENT

UK Biobank data can be obtained through https://www.ukbiobank.ac.uk/. Medication codes relevant to this study are available at https://github.com/YoungPKU‐SPS/statin_dementia. The R codes for the iCF algorithm are available at https://github.com/tianshengwang/iCF.