Key research priorities in methodological approaches for measuring the exposome and studying its role in the development of dementia

Abstract

There is growing recognition regarding the importance of the exposome, or the totality of exposures one experiences across the life course, in research on Alzheimer's disease and related dementias. However, the measurement of numerous exposures at once and over time, as well as modeling their effects on dementia risk, presents significant methodological challenges. Through community engagement and consensus‐building processes integrating input from multidisciplinary panels of experts, we identified critical priority topics for methods used in studying links between the exposome and dementia risk, along with advances needed to address those priorities. We identified nine priority topics: high‐dimensional and multimodal data, measurement error, harmonization across studies, mixtures of exposures, effect heterogeneity, exposure timing, cumulative exposures, reverse causation, and sample composition. This paper describes these priority topics and highlights areas where future research or the dissemination of existing methods could advance the state of existing science.

Highlights

Inherent complexities central to the measurement and modeling of the exposome and its relationship to dementia pose methodological challenges. We identified nine priority topics, such as measurement error, mixtures of exposures, and cumulative exposures. Modeling approaches should consider complexity but provide useful simplifications when possible. Investments in the development and dissemination of innovative approaches and methodological guidance are needed.

INTRODUCTION

The exposome is defined as the totality of exposures that individuals experience from conception to death and includes wide-ranging factors that influence how people live and age.^1–3 A key feature of exposome research is its focus on understanding the health impacts of multiple co-occurring exposures experienced across a person's life course. These exposures may be internal (e.g., inflammation, oxidative stress) or external (e.g., social, psychological, or environmental factors).¹ Consideration of this broad set of exposures requires careful attention to multiple temporal and cross-variable dimensions across which these co-occurring exposures influence and interact with each other to shape risk of later health outcomes such as Alzheimer's disease and related dementias (hereafter, dementia).

The exposome concept is aligned with increasing recognition that a wide range of environmental, socioeconomic, and lifestyle factors are relevant to the development of dementia. Research on the exposome encompasses investigations into the effects of these factors and combinations of these factors on dementia-related phenotypes, as well as investigations into interactions with genetic or biological determinants.^4,5 This integration of information on the broader social environment and health disparities in exposome research that also considers biological risk and biomarkers has been referred to as the compound exposome.⁶ Moreover, the focus on capturing exposure across the life course is in line with a growing literature on the importance of dementia risk factors across the life course and the potential for heterogeneous effects according to the period of life in which exposure occurred.^7–10 These findings, along with projected increases in the prevalence of dementia among older adults,¹¹ support the need for a shift from approaches that largely consider the effects of single exposures measured at a single point in time to an exposome-oriented approach.

Because the exposome is inherently complex, studying its effects on dementia raises important methodological challenges. The measurement of exposome features can itself be challenging, as available measurements may not always align well with constructs of interest or may include considerable error that lowers measurement precision and the power of subsequent analyses. Additionally, in exposome research, it is important to acknowledge that exposures do not occur in isolation but instead frequently co-occur and interact in complex ways that are often difficult to represent in traditional modeling frameworks. Furthermore, complexity across exposures also occurs alongside complexity over time and across the life course, adding important questions about the relevant timing of exposures during the life course, whether risk accumulates over time, and whether exposures at different time points interact with each other, among others. More traditional challenges in population studies, such as selection processes, remain relevant and should also be considered in the context of models primarily designed to handle other facets of complexity in exposome research.

The myriad methodological challenges in exposome research present important opportunities for the development of innovative methods and the dissemination of methodological guidance to promote best practices in the use of both novel and existing methodological approaches. Given the wide variety and large number of methodological challenges in exposome research, this paper aims to summarize and report on important priority topics for methodological development and guidance that could serve to advance population-based research on links between the exposome and dementia. While not exhaustive, we sought to identify key methodological priorities related to scientific areas critical to measuring and modeling the exposome and its impact on dementia. We include explanations of the selected methodological challenges, describe their importance to research on the links between the exposome and dementia, and highlight areas where innovative approaches are needed.

APPROACH

The Gateway Exposome Coordinating Center (GECC) is an interdisciplinary research center focused on fostering collaboration and building resources to facilitate research on the role of the exposome in dementia research. The priority topics identified in this paper are the result of a year-long process of community engagement and consensus building across multidisciplinary groups of experts, developed and organized by the GECC. The GECC conducted a series of six virtual town halls in the fall of 2024 to identify important broad gaps and priority topics for research on the exposome and its links with dementia. Attendees were recruited through direct outreach, with further dissemination conducted using flyers and mailing lists. A total of 382 scientists from 128 organizations in 21 countries attended the town halls, which were organized using the Open Space meeting format¹² with prompts asking stakeholders to reflect on questions such as what information they use to guide their work and what the key unanswered questions are in this area. Through this process, we identified 660 priority topics for research, ranging from novel exposures such as wildfires, novel measurement modalities such as wearables, or methodological considerations such as measurement error. While only 6% of the identified priority topics directly addressed methods or statistical approaches to analysis, methodological topics around measurement and modeling were pervasive throughout the town hall conversations. Other priority topics focused on substantive research areas (e.g., wildfires) will be discussed elsewhere.

Following the town hall meetings, identified methods-related priority topics were discussed among members of the GECC methods team, which included scientists across a range of disciplines, including epidemiology, biostatistics, environmental health, economics, and computer science. Following the development of an initial set of priorities, ideas were further discussed and refined among a larger group of experts convened at the GECC Gaps and Priorities Workshop in the spring of 2025, attended by 84 participants from 45 organizations and eight countries. Here, we highlight the nine priority topics selected at the end of this process: approaches to high-dimensional and multimodal data, measurement error, harmonization across studies, mixtures of exposures, effect heterogeneity, exposure timing, cumulative exposures, reverse causation, and sample composition (Box 1). We considered multiple criteria in topic selection, including (1) the importance of the methodological issue to research on the exposome, (2) areas of heightened relevance to dementia research, and (3) whether additional investments in methodological research and dissemination would have the potential to advance scientific progress (Table 1). Included topics relate to challenges across disciplines and in both the measurement (i.e., the assessment and characterization of the exposome) and modeling (i.e., designing studies and implementing data analysis approaches) of the exposome, though some topics have a greater focus on measurement challenges and some have a greater focus on modeling (Table 2). In the current paper, we focus on the presentation and discussion of priority topics, limitations of current approaches, and potential research ideas to further methodological research within each topic area; recommendations and solutions to the challenges raised here should be topics of future research. We include a glossary of technical terms in Box 2.

Box 1. A brief summary of nine priority topics for future methodological investment and research to improve the quality of research on the role of the exposome in the development of dementia

• High-dimensional and multimodal data: Rich data from a variety of sources, such as novel wearable sensors or linked administrative data, may provide important ways to improve exposure measurement, but appropriate methods are needed to handle high-dimensional and multimodal data.

• Measurement error: The difference between a measured or estimated value of a quantity and a true unknown value can affect one's ability to make inferences; better methods to quantify measurement error, evaluate potential bias due to measurement error, and incorporate measurement error into substantive analyses can improve the quality of study conclusions.

• Harmonization: Combining data sources and comparisons across different data resources needs to consider the harmonization of measures, and potentially the harmonization of methods and models as well, to ensure comparisons or pooled analyses are appropriate.

• Mixtures of exposures: Exposure mixtures are three or more co-occurring exposures; existing methods for pattern identification, toxic agent identification, subgroup estimation, overall effect estimation, and estimation of interactions are most commonly applied to chemical mixtures, though these methods may be more broadly applicable.

• Effect heterogeneity: There is effect heterogeneity when the effect of one exposure is different across levels of a second exposure or factor; theory-based and data-driven approaches to the identification of effect heterogeneity are complementary, though data-driven machine-learning approaches may have promise for evaluating many effect modifiers in a single analysis.

• Exposure timing: The comparison of effects across exposures experienced at different time points or life course periods is a core feature of exposome research; existing methods commonly used for environmental exposures (e.g., distributed lag models) may be more broadly useful, though considerations around methods for imputing or modeling life-course exposures and handling multicollinearity or differential measurement error over time deserve attention.

• Cumulative exposures: Quantification of how risk accumulates over time, capturing interactions between exposures at different time periods, can help guide the selection of candidate measures; comparisons between existing and innovative approaches to capture the accumulation of risk in exposure histories may be useful in matching estimates of cumulative exposures with theoretical reasoning or data-driven evidence on life course exposures.

• Reverse causation: Reverse causation exists when the observed association is because the outcome affects the exposure rather than vice versa; while analytic methods cannot definitively identify reverse causation, greater awareness of the potential source of bias is needed, and study design choices or sensitivity analyses can evaluate or reduce the risk of bias.

• Sample composition: Processes that govern the definition of the study sample can lead to challenges with internal validity (selection bias) or external validity (generalizability/ transportability) of estimates, though statistical approaches to account for selection bias, or to apply generalizability and transportability tools, can be used to mitigate concerns.

TABLE 1 Selection criteria motivating choice of nine priority topics for methodological research on the role of the exposome in the development of dementia.

Priority topic	Importance of exposome research	Importance to dementia research	Example value of additional investments
High-dimensional and multimodal data	Can improve existing measures of key exposome features.	In dementia research, reliance on exposure self-report can be problematic due to memory loss, motivating the use of objective measures or other sources of exposure information.	Methodological innovation and the dissemination of appropriate methods can facilitate the use of new data resources (e.g., novel wearable sensors or linked administrative data).
Measurement error	The large number of exposures measured over long time periods in exposome research complicates the consideration of measurement error and highlights its importance.	Measurement error in self-reported data is a greater concern in dementia research, given memory loss; measurement error in outcome ascertainment can also be a concern.	Consideration of how differential measurement error across exposures or across time impacts conclusions in exposome research is an important area for research.
Harmonization	Combining or comparing harmonized data sources can have important benefits to exposome research, including expanding the time range for longitudinal exposure measurement, including a larger number of exposome measures, or expanding the range of the exposure distribution.	Evidence suggests a wide range of different exposures and exposures across the life course are important to the development of dementia; combining data sources can improve exposure measurement and the quality of comparisons.	Dissemination of best practices in harmonization may help guide substantive researchers; methodological approaches to balance harmonization with confounder control in pooled studies could improve inference.
Mixtures of exposures	Assessing the effect of combinations of exposures is a key attribute of exposome research.	There is evidence on a wide range of modifiable risk factors for dementia, which often co-occur (e.g., cardio-metabolic risk factors).	The extension of mixture methods from chemical mixtures to broader applications could provide new insights; methodological adaptations and extensions may be necessary given new data types or study design features.
Effect heterogeneity	Effect heterogeneity enables the assessment of how the co-occurrence of risk factors impacts effect estimates, contributing to our understanding of co-occurring exposures.	Existing research has illustrated the existence of effect modification for many dementia risk factors, either by demographic characteristics or other risks.	Data-driven machine-learning methods to identify heterogeneous treatment effects in randomized controlled trials could be leveraged in observational exposome research to lend new insights given a large number of candidate effect modifiers.
Exposure timing	Consideration of exposures across the life course is a core feature of exposome research.	Dementia risk is shaped by exposures across the life course, including exposures in early life, midlife, and late life.	Evaluation and recommendations for the implementation of approaches to estimate exposure over long time periods could improve data resources; existing methodological approaches from environmental health may have broader applications in exposome research.
Cumulative exposures	Allows for consideration of how risk accumulates over time and how exposures interact across time periods.	Prior literature supports the theory of risk accumulation for many dementia risk factors.	Comparisons of existing or novel measures of cumulative exposures can help guide the selection of candidate measures.
Reverse causation	Consideration of multiple exposures across the life course raises the risk of reverse causation, particularly when comparing across time periods.	Given the long preclinical period in the development of dementia, reverse causation is a particularly salient issue in dementia research.	Dissemination of resources describing potential sensitivity analyses to evaluate reverse causation may increase awareness of this issue in exposome research.
Sample composition	Exposome research relying on selected samples with unique and detailed exposure data may suffer from concerns related to selection bias or lack of generalizability and transportability.	High mortality and study dropout among older adults at risk for dementia lead to a greater risk for selection bias in dementia research; many cohorts with dementia imaging or biomarker data are volunteer-based samples, and inferences, therefore, may have low external validity.	The development of resources on selection bias, transportability, and generalizability can help popularize existing approaches; methodological work to integrate solutions for selection bias into other methods popular in exposome research may be needed.

TABLE 2 Role of nine priority topic areas for methodological research in exposure assessment and measurement versus exposure-outcome modeling (i.e., study design and data analysis).

Priority topic	Connection to exposure assessment and measurement	Connection to exposure–outcome modeling (i.e., study design and data analysis)
High-dimensional and multimodal data	Can improve exposure measurement by integrating more data and incorporating diverse sources of information.	—
Measurement error	Reducing or accounting for measurement error may improve exposure measures.	Accounting for measurement error in analyses can reduce bias.
Harmonization	Exposure measures may require harmonization across data sources.	Harmonization can enable pooled analyses and models that may not be possible without multiple data sources.
Mixtures of exposures	Some mixture models focus on characterizing mixtures of exposures and refining exposure measurement.	Some mixture methods focus on quantifying associations between mixtures and outcomes of interest, as well as capturing interactions between mixture components.
Effect heterogeneity	—	Aims to capture variation in effect estimates across strata defined by different components of the exposome.
Exposure timing	Need to measure exposure across the life course.	Approaches seek to contrast effect estimates across different time periods.
Cumulative exposures	Need to measure exposure across the life course and summarize into a single summary of cumulative risk.	—
Reverse causation	—	Can lead to bias in outcome models without proper consideration and a preventive study design.
Sample composition	—	Can lead to bias in effect estimates or lack of generalizability to target population.

Box 2. Glossary of technical terms

• Analytic sample: The sample of participants included in a specific analysis.
• Critical period: A life course time period during which an exposure must occur to have an effect on a specific health outcome of interest.
• Exposome: The totality of exposures that individuals experience from conception to death, including wide-ranging factors that influence how people live and age.
• High-dimensional data: Data that include a large number of variables, which often serve as predictors or inputs for a model.
• Generalizability: The ability to extend inferences to a target population that includes the analytic sample and source population.
• Distributed lag model: A model with a large number of exposure periods, where the effect of different lags over time is modeled with a smooth function (e.g., a spline).
• Latent variable: A construct whose presence and value are inferred from observed indicators; for example, latent cognitive functioning can be inferred from scores on cognitive tests.
• Multimodal data: Data that combine information from different sources or modalities (e.g., sensor data and survey data).
• Multicollinearity: The presence of a high intercorrelation between independent variables included in a model.
• Multiple informant model: A single model to estimate the unconditional effects of exposures at different time windows, with formal statistical tests to assess differences between effects for different time windows.
• Selection bias: A difference between an estimated effect of interest and the true causal effect due to selection into the study sample.
• Sensitive period: A life course time period during which the effect of an exposure on a specific health outcome of interest is particularly strong.
• Supervised model: A type of prediction model that is estimated or trained using data that include both the predictors and outcomes of interest and then applied to new data to predict the same outcome observed in the training dataset.
• Synthetic cohort: A pooled dataset created by combining multiple individual cohorts through either stacking or linking the data.
• Target population: The population about which a researcher seeks to make inferences.
• Transportability: The ability to extend inferences to a target population that is partially or completely non-overlapping with the analytic sample and source population.
• Unsupervised model: A type of model that seeks to synthesize data on a set of predictors of interest without any available data on outcomes, with the intent to uncover patterns and structures in the data.

HIGH-DIMENSIONAL AND MULTIMODAL DATA

New data sources promise to complement more traditional modes of data collection and improve available measures of the exposome. Information from novel wearable sensors,¹³ linked administrative records,¹⁴ estimates of the physical environment from global models,¹⁵ biological signals of exposure from multi-omics data,¹⁶ and other sources of high-dimensional data provide exciting data resources that can be leveraged for scientific research. Combining these disparate data sources using frameworks that support multimodal data inputs may help improve exposure measurement. For example, measures of the physical environment (e.g., neighborhood disorder) based on multiple sources of high-dimensional data, such as Google Maps photos, reviews of neighborhood parks, satellite data, and population density predictions, may outperform simpler measures. However, despite the potential of these data to uncover novel insights, the use of high-dimensional and multimodal data for exposure measurement also brings new analytical challenges.^17,18 Combining larger quantities of high-dimensional data for measuring exposures can potentially improve measurement but also raises challenges in efforts to reduce dimensionality while retaining important signals in all available data. Multimodal approaches also seek to improve measurement by increasing the amount and diversifying the type of information used to quantify exposures, but they face distinct challenges related to aligning different sources of information with varying temporal and spatial resolutions.

Machine-learning approaches are well suited to address some of the challenges in using high-dimensional and multimodal data for exposure measurement and are increasingly being used for this purpose.¹⁹ Unsupervised machine models uncover patterns in high-dimensional data without reference to any external or gold-standard exposure measure and are useful when gold-standard exposure measures are not available or do not exist. However, unsupervised approaches are especially sensitive to the input data used in modeling, making replication across samples a challenge.²⁰ Additionally, both unsupervised and supervised (i.e., models to predict an exposure measure based on gold-standard data) models require large amounts of input data and, in the case of supervised machine-learning approaches, large amounts of labeled data (i.e., pairs of the gold standard exposure measurement with all corresponding input data) for model training.²¹ While it may be possible in some instances to apply trained models from data-rich to data-poor settings, model generalization across time and space is a concern. Additional research should seek to advance the use of machine-learning approaches to improve the measurement of exposome features by improving the generalization of predictive machine-learning models or improving approaches to adapt models to new contexts. Additionally, investments in resources to help substantive readers understand the strengths and limitations of these approaches will serve to improve their appropriate use in exposome research.

MEASUREMENT ERROR

Measurement error, defined as the difference between the measured or estimated value of a quantity and an unknown true value, can lead to bias in substantive analyses.²² Measurement error can be introduced via multiple different mechanisms, for example, through recall bias of self-report data, instrument error, limits of detection for chemicals in biospecimens, or prediction error in models to estimate environmental exposures (e.g., air pollution). Exposome research considering exposures over a long period of time may be particularly vulnerable to measurement error due to varying quality and quantity of historical exposure measurements and recall bias. Given potential bias due to measurement error, efforts to quantify measurement error, understand patterns of measurement error, and apply both novel and existing approaches to account for measurement error in analyses have the potential to improve the quality of statistical inferences.  

Substantial literature describes methods for understanding the potential magnitude and direction of expected bias due to measurement error based on the role of mismeasured constructs in analyses, the magnitude of the error, and the association between the error and other variables included in the analysis.^23,24 However, most of the literature focuses on only one or two exposures at a time,^25,26 whereas exposome research often considers a larger number of exposures simultaneously. Extensions of existing research and methods are needed to better characterize and estimate how measurement error may interact or compound across different exposures. Further, many exposome measures, such as air pollution and other features of the built environment, are also spatially patterned, bringing additional challenges associated with spatially correlated measurement error.^27,28 Characterizing the consequences of and correcting for different magnitudes and types of measurement errors (e.g., spatially correlated or not) across multiple exposome components is an important area for future methodological research. In tandem with methodological research extending existing measurement error approaches to handle unique characteristics of exposome research, the generation and dissemination of educational resources, application examples, and software implementations may also encourage the use of models that explicitly incorporate and consider measurement error in exposome research.

HARMONIZATION ACROSS STUDIES

Combining data from multiple studies can be useful for multiple purposes, such as increasing sample size, combining information across different parts of the life course,²⁹ and comparing settings with different exposure distributions. Given the expansive nature of exposome research, a single data source commonly has only some of the exposure information of interest for a given exposome research question. The limitations of a single data source, including the time period covered, the quality of exposure measures, or the range of the exposure distribution experienced by the study sample, can often be addressed by pooling data across studies. However, combining data or otherwise comparing across studies often requires harmonization, or efforts to derive comparable data for the purpose of joint or comparative analysis. Harmonization is seldom straightforward and often poses methodological challenges. For example, different studies may use different measures (e.g., different cognitive tests) for the same concept, and most analyses require some analytic procedure to make the data comparable. The specific research question, in combination with the comparability of data to be combined, will dictate how data can be combined and harmonized for use together.

When leveraging data from multiple studies, it is important to consider the capacity for harmonization of all relevant variables, including exposures and outcomes of interest, along with any confounders or effect modifiers. When the construct to be harmonized can be conceptualized as a latent variable and multiple indicators of the construct are measured, psychometric methods for statistical co-calibration can be useful in creating comparable measures across studies when the specific available indicators vary.^30,31 Though some overlap in indicators is typically needed to anchor comparisons, this approach, based on statistical co-calibration, is commonly used to harmonize cognitive functioning across studies on cognitive aging and dementia.^32,33 In other cases, it may be possible to use a calibration sample to harmonize different measures for the same underlying construct.³⁴ When harmonizing across contexts with vastly different age ranges, cultures, or languages, additional considerations are necessary.³⁵ Use of the same survey question does not necessarily imply successful harmonization when differences in translation, culture, or context may lead to differences in the interpretation of the survey question of interest.³⁶ Additionally, the harmonization of laboratory-based measures of chemicals or biomarkers raises challenges around differences in laboratory standards, assays, and protocols.^37,38 Dissemination of best practices for identifying and addressing harmonization issues would be useful to the broader research community, particularly for exposome research that incorporates cross-national and life course approaches. Additionally, potential differences in confounding structures across study populations (e.g., the social construct of race and associated racism may be a confounder in one setting, but the same construct may not exist in a different context) may also exist. The most appropriate methods to balance harmonization and control for confounding, if harmonization of all potential confounders is impossible, are an important area for future research.

MIXTURES OF EXPOSURES

The term mixture is often used to refer to three or more co-occurring exposures. Methods for modeling mixtures of exposures (different from statistical mixtures, as in reference 39) are diverse and prioritize different goals, such as pattern identification, toxic agent identification, estimation of effects in a priori defined subgroups, overall effect estimation, and estimation of interactions and non-linearities.⁴⁰ The development of methodological approaches focused on exposure mixtures has largely focused on chemical mixtures or mixtures of co-occurring air pollutants.^{41, 42} However, because measuring the broader exposome and its effects inherently requires studying multiple potentially highly correlated exposures simultaneously, mixture methods represent an opportunity to leverage existing methods for new research areas. Mixture methods address the aspects of exposome research that focus on understanding the ways in which different components of the exposome co-occur and estimating joint effects and interactions between co-occurring exposures in their effects on dementia phenotypes.

Recent efforts have sought to extend the concept of mixtures, applying them to the broader exposome measures such as social constructs around neighborhoods.⁴³ At the same time, some have cautioned against the application of mixture methods to diverse facets of the exposome (e.g., social isolation and environmental policy), arguing that mixtures of exposures would be most meaningful if they (1) had a common cause or shared mechanism, (2) informed a specific intervention, and (3) did not have clear directional relationships (i.e., one component causing another).⁴⁴ As the broader use of mixture methods grows, with further examples of more general applications beyond chemical mixtures, it will be important to continue to consider the theoretical limits of existing mixture methods for exposome research, leading to a better understanding of how different goals and subsequent interpretations of results may impact recommendations around appropriate usage. Additionally, because most mixture methods were developed with chemical mixtures in mind, the extension of mixture methods to broader exposome measures may require novel methodological extensions to existing approaches, to, for instance, accommodate binary or categorical exposure data. Other areas where research could enhance the use of mixture methods in exposome research include ensuring compatibility with different study designs, developing extensions to existing methods to allow for the use of sampling weights, considering differential measurement error across mixture components, and considering bias amplification in settings with highly correlated exposures.⁴⁵

EFFECT HETEROGENEITY

Even when groups of exposures should not conceptually be considered mixtures, it is still of interest to understand the effects of co-occurring exposures in exposome research. Characterizing effect heterogeneity (i.e., when the effect of one exposure is different across levels of a second exposure or other factor) provides another avenue for studying the impacts of multiple exposures. For example, evidence suggests that the effect of exposure to diabetes on dementia risk is larger among those with hypertension than among those without hypertension, illustrating the importance of understanding joint exposures to different groups of risk factors.⁴⁶ Understanding this form of effect heterogeneity may serve multiple purposes. Insights into heterogeneous effects could lead to conclusions about the potential variability in effect sizes for groups with differing vulnerability or susceptibility, help move toward a so-called precision medicine approach by identifying potential interventions with large effects for specific subpopulations, or yield insights into biological mechanisms and mechanistic interaction between different exposures. Furthermore, in tandem with research on differential exposure, understanding effect heterogeneity across diverse factors related to the social environment may help identify potential mechanisms that may underlie observed health disparities.⁴⁷

In exposome research with many exposures, traditional approaches based on adding interaction terms to regression models are likely not scalable. Instead, one option is to apply data-driven machine-learning-based approaches to identify heterogeneity across a wide range of factors.^48,49 Though these methods have been most commonly utilized to identify heterogeneous treatment effects in the setting of randomized controlled trials,⁵⁰ they can be applied to observational research as well. Additional efforts should be directed toward better understanding and disseminating best practices for the use of such data-driven effect heterogeneity methods in exposome research and the most appropriate ways to pair these data-driven approaches with hypothesis-driven models based on theory and initial evidence from exploratory research.

EXPOSURE TIMING

In addition to the consideration of multiple exposures occurring at once, the exposome also introduces complexity when one considers dementia risk factors across the life course. Evaluating heterogeneity in exposure–outcome associations by exposure timing (e.g., age at exposure or time relative to outcome) has important implications for both understanding disease biology and planning effective interventions. Differences may also shed light on existing heterogeneity across the literature if exposure timing is not taken into consideration in comparisons between studies. For example, prior work reported differences in the effect of high educational attainment achieved early in life versus later in life on late-life cognition,⁵¹ suggesting that differences in the effect of education on cognition in two populations with differences in the timing of educational attainment could be explained by the heterogeneity in exposure timing. The most relevant life course stage for dementia risk may vary by risk factor, and research on exposure timing can provide quantitative evidence to further our understanding of critical or sensitive periods for life course dementia risk factors.

Designing quantitative assessments of exposure timing requires consideration of multiple methodological issues related to both measurement and modeling, including challenges around the collection of data or estimation of exposures over long time periods, the identification of specific key exposure periods, and the comparison of effects across exposure timings. Despite growing data resources for life course research, a lack of available exposure data covering a wide age range often remains a challenge. Approaches to reconstruct exposure histories, including linkage of historical records, environmental records of the physical environment, or the construction of exposure histories via prospective or retrospective reporting in survey data, all have their own set of advantages and limitations. Alternatively, linkage across cohorts can allow researchers to model exposure across the full time span of interest using information from other data sources via a synthetic cohort design.^52,53 For chemical exposures, variability in the half-life of different chemicals measured in different tissues may impact the number and timing of measures needed to construct a longitudinal history of exposure.

Some prior research on reconstructing exposure histories using survey data focused on multivariate record-level imputation of missingness in studies that sparsely cover the time period of interest for exposure measurement.^54,55 Alternatively, trajectory-based modeling approaches could also be used either to impute record-level missingness or to provide estimates of exposure with more granular time intervals than existing study intervals. Research is needed to assess the sensitivity of approaches for modeling missing exposure data to a range of analytic decisions and to provide frameworks for evaluating the quality of predictive models for exposure histories.

Analytic approaches to identifying key exposure periods and evaluating effect heterogeneity by exposure timing, including distributed lag models (DLMs)⁵⁶ or multiple informant models,⁵⁷ have typically been used in environmental health applications focused on pollution or chemical exposures but may be applied to other topic areas. Dissemination efforts and application examples may be useful in illustrating the potential benefit of these approaches in broader dementia exposome research. Quantitative comparisons between approaches may also contribute more concrete guidance on the most appropriate approach given a specific research question and dataset. Furthermore, many available methods do not explicitly consider or address important challenges with time-series data. For example, challenges such as multicollinearity in repeated exposure data or the potential for differential measurement error over time can impact the ability of models to identify key exposure periods or compare effects across different exposure timings and deserve future attention.

CUMULATIVE EXPOSURES

In addition to understanding the heterogeneity of effects by exposure timing, exposome research also places an emphasis on the importance of capturing how risk develops over time. While comparisons of exposure timing do not consider the ways in which exposures at different time periods interact, cumulative exposure measures attempt to capture these interactions by considering the accumulation of risk across the life course. Cumulative measures should, at a minimum, consider the duration of time spent at different exposure levels. An example of this type of quantification is the calculation of pack years of exposure to smoking, which integrates information on the number of packs of cigarettes smoked per day and the number of years smoked.⁵⁸ While providing advantages over measures of current smoking only, these approaches still fail to capture differences between, for example, someone who smoked 10 packs a day for 1 year and four packs a day the next year versus someone who smoked seven packs a day for 2 years in a row. Threshold effects or effects of specific patterns or trajectories of exposure will be missed by common approaches, including the use of exposure unit years (e.g., pack years). When sufficient evidence exists, evidence-based theoretical decisions should guide choices around summaries of cumulative exposure. However, when such evidence does not exist, data-driven comparisons of candidate measures may be useful. Regardless, throughout the literature, greater attention is needed to the specific life course hypotheses being tested, given specific approaches for estimating cumulative exposures.

Additionally, new methods of estimating cumulative exposures based on exposure over time could potentially be adapted from existing approaches for summarizing risk across other dimensions in different settings. For example, polygenic risk scores for genetic data consider many exposures over the genome, and similar methods could be used to consider many exposures over time. In exposome research considering multiple risk factors simultaneously, it will also be important to consider how heterogeneity in approaches to quantifying cumulative exposures (e.g., total number of days exposed to a heat wave versus cumulative lead exposure measured via bone lead) may impact comparisons or efforts to summarize joint exposure distributions.

REVERSE CAUSATION

When the temporal ordering of the exposure and outcome is unclear, observed associations between a hypothesized exposure and outcome can be due to reverse causation or the effect of the outcome on the exposure, rather than vice versa. In many cases, causation may occur in both directions (i.e., reciprocal causation), biasing effect estimates specific to either direction. Reverse causation is particularly important to consider in exposome research, considering a lifetime of exposures across different ages or time periods. For example, when comparing across effects by exposure timing, it is important to consider whether observed heterogeneity may be attributable to differences in the impact of reverse causation on findings (i.e., effects of exposure later in life and therefore closer in time to outcome ascertainment may be more susceptible to reverse causation), rather than true differences in effects by exposure timing. Lack of clarity on the potential for reverse causation may also affect decisions around the best way to quantify cumulative exposures and whether to exclude exposure measurements more proximal to outcome ascertainment. Disentangling the impact of reverse causation can be even more challenging in scenarios where insufficient follow-up is available, leading to uncertainty in establishing temporality between the exposure and outcome.

Challenges related to reverse causation can be particularly salient in dementia research, given the long preclinical period of the disease.⁵⁹ For example, observed associations between dementia and lower body mass index (BMI) at older ages are believed to reflect increasing frailty as a result of the disease process, as opposed to low BMI increasing the risk of dementia.^60,61 Since these effects would not appear among younger people, this can create the appearance of heterogeneity in the effects of BMI by age when none exists. Therefore, when evaluating the effects of risk factors across the life course in exposome research, it is important to consider whether heterogeneity across age may be due to true effect heterogeneity or reverse causation. Although analytic methods cannot definitively identify reverse or reciprocal causation, a range of different study design choices or sensitivity analyses can help researchers alleviate the risk of bias resulting from reverse causation or evaluate the potential for reverse causation to influence results. For example, excluding participants with preclinical disease at baseline has the potential to limit risk for reverse causation, though doing so changes the study sample and limits the potential to observe effects on early disease, especially in studies with shorter follow-up. Dissemination of resources on best practices for evaluating reverse causation in research on the dementia exposome may serve to increase awareness and evaluation of how reverse causation can impact observational research findings.

SAMPLE COMPOSITION

Typically, the characteristics of participants in an analytic sample (i.e., the study sample) are not identical to the characteristics of the group that a researcher seeks to make inferences about (i.e., the target population).^62,63 The selection of the analytic sample can compromise inferences for the included analytic sample (internal validity), inferences for the target population (external validity), or both. We use the term selection bias to refer to selection processes that lead to a lack of internal validity and the term lack of generalizability/transportability to refer to selection processes that lead to a lack of external validity. Selection bias commonly arises when selection into an analytic sample due to consent, mortality, or other forms of attrition is jointly related to both the exposome component of interest and the outcome of interest (e.g., dementia or cognitive decline) in a given study.^62,64,65 In dementia research, selection bias due to frailty or high mortality among older adults at risk of dementia is often of particular concern (e.g., in studies of smoking and cognitive decline, where the conclusion changes after accounting for selection bias⁶⁶). Additionally, conclusions regarding how the social environment might interact with other aspects of the exposome to explain health disparities may be incorrect if minority groups are less likely to consent to be in a study and are more likely to drop out.⁶⁷

Even when there is no selection bias, it is still possible that differences between the analytic sample and the target population make it such that findings from a given study are inaccurate for a target population of interest. Generalizability or transportability methods focus on reconciling differences between the analytic sample and target populations, when the analytic sample is either a subset of the target population (generalizability) or when the analytic sample is instead partially or completely non-overlapping with the target population (transportability).^68,69 In applying generalizability or transportability methods, it is important to pay attention to the distributions of the exposure, effect modifiers, confounders, and mediators for a specific study question.

For example, an exposome study may recruit participants from the Washington, D.C., metro area aged 70 to 74 years old and require them to annually wear accelerometers for 3 weeks in addition to undergoing cognitive testing to measure cognitive decline. In an ideal world, the study would have 100% consent, 0% mortality, and no attrition due to other factors. However, the exclusion of eligible participants due to lack of consent, attrition, or mortality may all contribute to selection bias. Further, researchers may seek to use these data to draw conclusions about the US population more generally, and differences in the distributions of the exposure, effect modifiers, confounders, and mediators between the analytic sample and the broader US population may impact the transportability of the estimates.

Though selection bias and concerns around transportability and generalizability have the potential to impact all health research, these issues may be particularly relevant for studies of the exposome, where selection due to linkage with detailed exposome data may be a substantial concern. For example, studies requiring detailed exposure measures from ecologic momentary assessments⁷⁰ may be more susceptible to differential selection processes due to lack of consent or participant demands. Though there is a vast literature on methods to account for selection bias either at baseline or during follow-up^62,64,65 and the use of these approaches is somewhat common in the epidemiologic literature, these approaches, including inverse probability weighting,⁶⁶ are less commonly integrated into exposome research. A lack of compatibility between some methods designed to focus on other areas of exposome research (e.g., mixture methods for multiple exposures) and approaches to account for selection bias highlights important areas for research. There is a smaller literature on generalizability and transportability methods (e.g., inverse odds weighting) and, therefore, a greater need for further dissemination across disciplines with applied examples focused on meaningful use cases.^71,72 Additionally, while some existing literature illustrates the performance of these approaches in simulation studies,^73–75 there is further need for additional research on the performance of generalizability and transportability methods across a wider range of realistic scenarios where the assumptions that underpin these methods are not fully met.

CONCLUSION

Across the nine priority topics highlighted in this paper, there are many exciting and important opportunities for efforts focused on methodological dissemination and methodological research to improve the quality of evidence on the dementia exposome. Accurately representing and modeling the complexity associated with the exposome is an overarching challenge across existing approaches and topic areas. Many approaches focus on one area of complexity at a time (e.g., temporal complexity or co-occurring exposures), but future research is needed to combine features across available approaches to conduct, for example, life course research on mixtures of exposures. In addition to highlighting areas for novel methods development, it is also important to consider how methodological approaches typically applied in specific disciplines or contexts may have broader use cases in multidisciplinary exposome research. Given existing disciplinary silos, the appropriate use of analytic methods in exposome research could be improved through cross-fertilization across disciplines and the development of open-access resources to effectively translate and disseminate information on available tools and approaches across fields.

Despite the focus on considering realistic complexity in dementia exposome research, methodological guidance should also consider the utility and added value of complexity. The most useful research necessarily relies on simplifications of complex environments, balancing data constraints, method limitations, and the interpretability of conclusions. It may not be possible to address all possible sources of bias in every situation, but methodological research and guidance can help substantive researchers identify and address the specific biases most likely to have the largest impacts on results and conclusions for a given research question and study design. Focus on methodological research in the key areas highlighted throughout the current paper will facilitate the development of innovative approaches, while also guiding researchers in the appropriate use and consideration of methodological issues. Ultimately, by addressing key limitations of existing approaches and improving the appropriate use of available methods, investments in key priority areas for methodological research have the potential to improve the quality of existing research and kickstart new efforts to uncover the role of the exposome in the development of dementia.

ACKNOWLEDGMENTS

National Institutes of Health/National Institute on Aging (U24AG088894 to J.L. and S.D.A.; R01AG030153 to J.L. and S.D.A.; R00AG075317 to E.H.L.; R00AG084769 to K.L.K.; R01AG065359 to J.W. (Weuve), R01AG067497 to J.W. (Weuve), R13AG064971 to J.W. (Weuve).

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

This review did not include human subject research, so obtaining consent was not necessary.

References

Wild CP. The exposome: from concept to utility. Int J Epidemiol. 2012;41:24‐32. doi:10.1093/ije/dyr236

Wild CP. Complementing the genome with an “exposome”: the outstanding challenge of environmental exposure measurement in molecular epidemiology. Cancer Epidemiol Biomarkers Prev. 2005;14:1847‐1850. doi:10.1158/1055‐9965.EPI‐05‐0456

Miller GW. The exposome: a new paradigm for the environment and health. Academic Press; 2020.

Livingston G, Huntley J, Liu KY, et al. Dementia prevention, intervention, and care: 2024 report of the Lancet standing Commission. The Lancet. 2024;404:572‐628. doi:10.1016/S0140‐6736(24)01296‐0

Migliore L, Coppedè F. Gene–environment interactions in Alzheimer disease: the emerging role of epigenetics. Nat Rev Neurol. 2022;18:643‐660. doi:10.1038/s41582‐022‐00714‐w

Nwanaji‐Enwerem JC, Jackson CL, Ottinger MA, et al. Adopting a “compound” exposome approach in environmental aging biomarker research: a call to action for advancing racial health equity. Environ Health Perspect. 2021;129:045001. doi:10.1289/EHP8392

Yang W, Li X, Pan K‐Y, et al. Association of life‐course depression with the risk of dementia in late life: a nationwide twin study. Alzheimers Dement. 2021;17:1383‐1390. doi:10.1002/alz.12303

Krishna M, Krishnaveni GV, Sargur V, et al. Size at birth, life course factors, and cognitive function in late life: findings from the MYsore study of Natal effects on Ageing and Health (MYNAH) cohort in South India. Int Psychogeriatr. 2022;34:353‐366. doi:10.1017/S1041610221001186

George KM, Lutsey PL, Kucharska‐Newton A, et al. Life‐course individual and neighborhood socioeconomic status and risk of dementia in the atherosclerosis risk in communities neurocognitive study. Am J Epidemiol. 2020;189:1134‐1142. doi:10.1093/aje/kwaa072

Marden JR, Tchetgen Tchetgen EJ, Kawachi I, Glymour MM. Contribution of socioeconomic status at 3 life‐course periods to late‐life memory function and decline: early and late predictors of dementia risk. Am J Epidemiol. 2017;186:805‐814. doi:10.1093/aje/kwx155

Nichols E, Steinmetz JD, Vollset SE, et al. Estimation of the global prevalence of dementia in 2019 and forecasted prevalence in 2050: an analysis for the Global Burden of Disease Study 2019. Lancet Public Health. 2022;7:e105‐e125. doi:10.1016/S2468‐2667(21)00249‐8

Owen H. Open Space Technology: A User's Guide. Berrett‐Koehler Publishers; 2008.

Virginia Anikwe C, Friday Nweke H, Chukwu Ikegwu A, et al. Mobile and wearable sensors for data‐driven health monitoring system: state‐of‐the‐art and future prospect. Expert Syst Appl. 2022;202:117362. doi:10.1016/j.eswa.2022.117362

Jutte DP, Roos LL, Brownell MD. Administrative record linkage as a tool for public health research. Annu Rev Public Health. 2011;32:91‐108. doi:10.1146/annurev‐publhealth‐031210‐100700

Shaddick G, Thomas ML, Green A, et al. Data integration model for air quality: a hierarchical approach to the global estimation of exposures to ambient air pollution. J R Stat Soc Ser C Appl Stat. 2018;67:231‐253. doi:10.1111/rssc.12227

Walker DI, Valvi D, Rothman N, Lan Q, Miller GW, Jones DP. The metabolome: a key measure for exposome research in epidemiology. Curr Epidemiol Rep. 2019;6:93‐103. https://doi.org/10.1007/s40471‐019‐00187‐4

Lahat D, Adali T, Jutten C. Multimodal data fusion: an overview of methods, challenges, and prospects. Proc IEEE. 2015;103:1449‐1477. doi:10.1109/JPROC.2015.2460697

Johnstone IM, Titterington DM. Statistical challenges of high‐dimensional data. Philos Trans R Soc A. 2009;367:4237‐4253. doi:10.1098/rsta.2009.0159

Ravì D, Wong C, Deligianni F, Berthelot M, Andreu‐Perez J, Lo B, et al. Deep learning for health informatics. IEEE J Biomed Health Inf. 2017;21:4‐21. doi:10.1109/JBHI.2016.2636665

Liu T, Yu H, Blair RH. Stability estimation for unsupervised clustering: a review. Wiley Interdiscip Rev Comput Stat. 2022;14:e1575. doi:10.1002/wics.1575

Ghassemi M, Naumann T, Schulam P, Beam AL, Chen IY, Ranganath R. A review of challenges and opportunities in machine learning for health. AMIA Jt Summits Transl Sci Proc. 2020;2020:191‐200.

Rothman KJ, Greenland S, Lash TL. Modern epidemiology. Wolters Kluwer Health/Lippincott Williams & Wilkins Philadelphia; 2008.

Schennach SM. Recent advances in the measurement error literature. Annu Rev Econ. 2016;8:341‐377. doi:10.1146/annurev‐economics‐080315‐015058

Innes GK, Bhondoekhan F, Lau B, Gross AL, Ng DK, Abraham AG. The measurement error elephant in the room: challenges and solutions to measurement error in epidemiology. Epidemiol Rev. 2021;43:94‐105. doi:10.1093/epirev/mxab011

Lash TL, Fox MP, MacLehose RF, Maldonado G, McCandless LC, Greenland S. Good practices for quantitative bias analysis. Int J Epidemiol. 2014;43:1969‐1985. doi:10.1093/ije/dyu149

Petersen JM, Ranker LR, Barnard‐Mayers R, MacLehose RF, Fox MP. A systematic review of quantitative bias analysis applied to epidemiological research. Int J Epidemiol. 2021;50:1708‐1730. doi:10.1093/ije/dyab061

Sheppard L, Burnett RT, Szpiro AA, et al. Confounding and exposure measurement error in air pollution epidemiology. Air Qual Atmos Health. 2012;5:203‐216. doi:10.1007/s11869‐011‐0140‐9

Keller JP, Chang HH, Strickland MJ, Szpiro AA. Measurement error correction for predicted spatiotemporal air pollution exposures. Epidemiology. 2017;28:338. doi:10.1097/EDE.0000000000000623

Kezios KL, Glymour MM, Zeki Al Hazzouri A. An introduction to longitudinal synthetic cohorts for studying the life course drivers of health outcomes and inequalities in older age. Curr Epidemiol Rep. 2024;12:2. doi:10.1007/s40471‐024‐00355‐1

Chan KS, Gross AL, Pezzin LE, Brandt J, Kasper JD. Harmonizing measures of cognitive performance across international surveys of aging using item response theory. J Aging Health. 2015;27:1392‐1414. doi:10.1177/0898264315583054

Jacobson LP, Parker CB, Cella D, Mroczek DK, Lester BM. Approaches to protocol standardization and data harmonization in the ECHO‐wide cohort study. Pediatr Res. 2024;95:1726‐1733. doi:10.1038/s41390‐024‐03039‐0

Gross AL, Li C, Briceño EM, et al. Harmonisation of later‐life cognitive function across national contexts: results from the harmonized cognitive assessment protocols. Lancet Healthy Longev. 2023;4:e573‐e583. doi:10.1016/S2666‐7568(23)00170‐8

Mukherjee S, Choi S‐E, Lee ML, et al.; Link to external site this link will open in a new window. Cognitive domain harmonization and cocalibration in studies of older adults. Neuropsychology. 2023;37:409‐423. doi:10.1037/neu0000835

Wu Y, Hayes‐Larson E, Zhou Y, et al. Statistical harmonization of versions of measures across studies using external data: self‐rated health and self‐rated memory. Ann Epidemiol. 2025;102:86‐90. doi:10.1016/j.annepidem.2025.01.002

Kobayashi LC, Jones RN, Briceño EM, et al. Cross‐national comparisons of later‐life cognitive function using data from the Harmonized Cognitive Assessment Protocol (HCAP): considerations and recommended best practices. Alzheimers Dementia. 2024;20:2273‐2281. doi:10.1002/alz.13694

Beaton DE, Bombardier C, Guillemin F, Ferraz MB. Guidelines for the process of cross‐cultural adaptation of self‐report measures. Spine. 2000;25:3186‐3191. doi:10.1097/00007632‐200012150‐00014

Iqbal K, Intemann T, Börnhorst C, Zhang J, Aleksandrova K. Approaches for harmonization of biomarker data from multiple studies: a narrative methodological review. AJE Adv. 2025;1:uuaf005. doi:10.1093/ajeadv/uuaf005

Kannan K, Stathis A, Mazzella MJ, et al. Quality assurance and harmonization for targeted biomonitoring measurements of environmental organic chemicals across the children's health exposure analysis resource laboratory network. Int J Hyg Environ Health. 2021;234:113741. doi:10.1016/j.ijheh.2021.113741

Compiani G, Kitamura Y. Using mixtures in econometric models: a brief review and some new results. The Econometrics Journal. 2016;19:C95‐127. doi:10.1111/ectj.12068

Joubert BR, Kioumourtzoglou M‐A, Chamberlain T, et al. Powering research through innovative Methods for Mixtures in Epidemiology (PRIME) Program: novel and Expanded Statistical Methods. Int J Environ Res Public Health. 2022;19:1378. doi:10.3390/ijerph19031378

Dominici F, Peng RD, Barr CD, Bell ML. Protecting human health from air pollution: shifting from a single‐pollutant to a multipollutant approach. Epidemiology. 2010;21:187‐194. doi:10.1097/EDE.0

Carpenter DO, Arcaro K, Spink DC. Understanding the human health effects of chemical mixtures. Environ Health Perspect. 2002;110:25‐42. doi:10.1289/ehp.02110s125

Bather JR, Robinson TJ, Goodman MS. Bayesian kernel machine regression for social epidemiologic research. Epidemiology. 2024;35:735‐747. doi:10.1097/EDE.0000000000001777

Schnake‐Mahl A, Hamra GB. Mixture models for social epidemiology: opportunities and cautions. Epidemiology. 2024;35:748‐752. doi:10.1097/EDE.0000000000001778

Weisskopf MG, Seals RM, Webster TF. Bias amplification in epidemiologic analysis of exposure to mixtures. Environ Health Perspect. 2018;126:047003. doi:10.1289/EHP2450

Xu WL, Qiu CX, Wahlin Å, Winblad B, Fratiglioni L. Diabetes mellitus and risk of dementia in the Kungsholmen project. Neurology. 2004;63:1181‐1186. doi:10.1212/01.WNL.0000140291.86406.D1

Ward JB, Gartner DR, Keyes KM, Fliss MD, McClure ES, Robinson WR. How do we assess a racial disparity in health? Distribution, interaction, and interpretation in epidemiological studies. Ann Epidemiol. 2019;29:1‐7. doi:10.1016/j.annepidem.2018.09.007

Wager S, Athey S. Estimation and inference of heterogeneous treatment effects using random forests. J Am Statist Assoc. 2018;113:1228‐1242. doi:10.1080/01621459.2017.1319839

Künzel SR, Sekhon JS, Bickel PJ, Yu B. Metalearners for estimating heterogeneous treatment effects using machine learning. Proc Natl Acad Sci. 2019;116:4156‐4165. doi:10.1073/pnas.1804597116

Inoue K, Adomi M, Efthimiou O, et al. Machine learning approaches to evaluate heterogeneous treatment effects in randomized controlled trials: a scoping review. J Clin Epidemiol. 2024;176:111538. doi:10.1016/j.jclinepi.2024.111538

Soh Y, Whitmer RA, Mayeda ER, et al. Timing and level of educational attainment and late‐life cognition in the KHANDLE study. Alzheimers Dement. 2024;20:593‐600. doi:10.1002/alz.13475

Hughes RA, Tilling K, Lawlor DA. Combining longitudinal data from different cohorts to examine the life‐course trajectory. Am J Epidemiol. 2021;190:2680‐2689. doi:10.1093/aje/kwab190

Kezios KL, Zimmerman SC, Buto PT, et al. Overcoming data gaps in life course epidemiology by matching across cohorts. Epidemiology. 2024;35:610. doi:10.1097/EDE.0000000000001761

Shaw C, Wu Y, Zimmerman SC, et al. Comparison of imputation strategies for incomplete longitudinal data in life‐course epidemiology. Am J Epidemiol. 2023;192:2075‐2084. doi:10.1093/aje/kwad139

Huque MH, Carlin JB, Simpson JA, Lee KJ. A comparison of multiple imputation methods for missing data in longitudinal studies. BMC Med Res Methodol. 2018;18:1‐16. doi:10.1186/s12874‐018‐0615‐6

Gasparrini A. Modeling exposure–lag–response associations with distributed lag non‐linear models. Stat Med. 2014;33:881‐899. doi:10.1002/sim.5963

Sánchez BN, Hu H, Litman HJ, Téllez‐Rojo MM. Statistical methods to study timing of vulnerability with sparsely sampled data on environmental toxicants. Environ Health Perspect. 2011;119:409‐415. doi:10.1289/ehp.1002453

Leffondré K, Abrahamowicz M, Siemiatycki J, Rachet B. Modeling smoking history: a comparison of different approaches. Am J Epidemiol. 2002;156:813‐823. doi:10.1093/aje/kwf122

Jack CR, Knopman DS, Jagust WJ, et al. Hypothetical model of dynamic biomarkers of the Alzheimer's pathological cascade. Lancet Neurol. 2010;9:119‐128. doi:10.1016/S1474‐4422(09)70299‐6

Suemoto CK, Gilsanz P, Mayeda ER, Glymour MM. Body mass index and cognitive function: the potential for reverse causation. Int J Obes. 2015;39:1383‐1389. doi:10.1038/ijo.2015.83

Pedditizi E, Peters R, Beckett N. The risk of overweight/obesity in mid‐life and late life for the development of dementia: a systematic review and meta‐analysis of longitudinal studies. Age Ageing. 2016;45:14‐21. doi:10.1093/ageing/afv151

Lu H, Cole SR, Howe CJ, Westreich D. Toward a clearer definition of selection bias when estimating causal effects. Epidemiology. 2022;33:699‐706. doi:10.1097/EDE.0000000000001516

Lash TL, VanderWeele TJ, Haneause S, Rothman K. Modern Epidemiology. Lippincott Williams & Wilkins; 2020.

Hernán MA, Hernández‐Díaz S, Robins JM, A Structural Approach to Selection Bias : Epidemiology n.d.

Infante‐Rivard C, Cusson A. Reflection on modern methods: selection bias—a review of recent developments. Int J Epidemiol. 2018;47:1714‐1722. doi:10.1093/ije/dyy138

Weuve J, Tchetgen Tchetgen EJ, Glymour MM, et al. Accounting for bias due to selective attrition: the example of smoking and cognitive decline. Epidemiology. 2012;23:119‐128. doi:10.1097/EDE.0b013e318230e861

Gilmore‐Bykovskyi AL, Jin Y, Gleason C, et al. Recruitment and retention of underrepresented populations in Alzheimer's disease research: a systematic review. Alzheimers Dement. 2019;5:751‐770. doi:10.1016/j.trci.2019.09.018

Hayes‐Larson E, Zhou Y, Rojas‐Saunero LP, et al. Methods for extending inferences from observational studies: considering causal structures, identification assumptions, and estimators. Epidemiology. 2024;35:753. doi:10.1097/EDE.0000000000001780

Degtiar I, Rose S. A review of generalizability and transportability. Annu Rev Stat Appl. 2023;10:501‐524. doi:10.1146/annurev‐statistics‐042522‐103837

Shiffman S, Stone AA, Hufford MR. Ecological momentary assessment. Annu Rev Clin Psychol. 2008;4:1‐32. doi:10.1146/annurev.clinpsy.3.022806.091415

Vuong Q, Metcalfe RK, Ling A, Ackerman B, Inoue K, Park JJ. Systematic review of applied transportability and generalizability analyses: a landscape analysis. Ann Epidemiol. 2025;104:61‐70. doi:10.1016/j.annepidem.2025.03.001

Levy NS, Arena PJ, Jemielita T, et al. Use of transportability methods for real‐world evidence generation: a review of current applications. J Comp Eff Res. 2024;13:e240064. doi:10.57264/cer‐2024‐0064

Schmid I, Rudolph KE, Nguyen TQ, et al. Comparing the performance of statistical methods that generalize effect estimates from randomized controlled trials to much larger target populations. Commun Stat Simul Comput. 2022;51:4326‐4348. doi:10.1080/03610918.2020.1741621

Kern HL, Stuart EA, Hill J, Green DP. Assessing methods for generalizing experimental impact estimates to target populations. J Res Educ Eff. 2016;9:103‐127. doi:10.1080/19345747.2015.1060282

Rudolph KE, Díaz I, Rosenblum M, Stuart EA. Estimating population treatment effects from a survey subsample. Am J Epidemiol. 2014;180:737‐748. doi:10.1093/aje/kwu197