Attrition and Generalizability of Cognitive Aging Studies - A Population-Based Perspective
The number of elderly people suffering from cognitive impairment is expected to increase substantially along the population aging. In addition to the health concern for those affected, this may also cause a greater demand on our health care systems. Longitudinal studies where individuals are followed over a long time period on their cognitive performance have become central for studying cognitive aging and identifying modifiable risk and preventive factors. However, such studies face a difficulty from the fact that especially older participants tend to decline continued participation. Ignoring attrition may result in an overestimation of cognitive performance and moreover bias the contribution of risk factors on cognitive aging. While implications and adjustment for attrition on estimates of cognitive function is an active filed of research, selection before assessment is less well studied. Although determining the extent and consequences of different inclusion mechanisms are crucial for establishing generalizability of results beyond the study sample. In the proposed project we will identify possible inclusion mechanism and study how estimates of cognitive performance are influenced by attrition prior to assessment. The combination of data from a longitudinal study on cognition, health, and aging (the Betula study) and Swedish register data will provide unique opportunities to study possible inclusion mechanisms and their implications.
Final report
The aim of the project was to study how different inclusion and dropout mechanisms affect estimates in studies of cognitive aging. The project has resulted in 3 articles, both method and empirical, and software in the form of an R-package.
In Josefsson and Daniels (2021), the purpose was first, to study the effect of becoming a widow / widower on episodic memory and second, to develop a better method for investigating this. We used longitudinal data from the Betula study, which allowed us to follow individuals over time on both widow status and memory. We included people who were married at baseline and then compared memory performance for individuals who became widows to those who were still married. The results did not show a significant difference between the groups. We used G-computation, which is a method for estimating a (causal) effect of a variable that varies over time (in our case widowhood status). Our method, unlike previous studies, can adjust for selective survival and non-ignorable dropout and implements Bayesian machine learning instead of standard regression models. The results underline the importance of longitudinal data for studying the effect of risk factors on cognitive aging, and that machine learning is an effective and simple tool for estimating these models.
In Josefsson et al. (2021) the aim was to estimate cognitive aging over the life span. The validity of previous longitudinal studies is unclear since i) the samples are often not representative of the population, ii) they did not take into account selective participation, and iii) did not control for practice effects. Here we combined cohort data for the sample with register data for both the sample and the target population to adjust for selective participation, and further extended the method in Study 1 to adjust for non-random dropout, death and practice effects. The results show that cognitive performance is overestimated with previous methods but the change over the life span is quite similar, i.e. we see that the memory is largely unchanged until about 60 years of age and a sharp decrease in episodic memory after that.
In Farnsworth von Cederwald et al. (2022) the aim was to investigate whether changes in cardiovascular risk profiles over time predict dementia and/or episodic memory impairment. As a measure of cardiovascular risk, a composite measure was used based on systolic blood pressure, diagnosis of hypertension, BMI, diabetes, smoking, age and gender. Three groups were identified, groups with stable, average and impaired cardiovascular health. Modeling was done with Bayesian machine learning for survival data with competing events (ie dementia and death). The results showed that individuals with impaired cardiovascular health were at significantly higher risk for both vascular dementia and Alzheimer's disease at older ages; but also an increased risk of episodic memory impairment already in middle age. For carriers of the risk gene Apoe E4, a maintained cardiovascular health can significantly reduce the risk of dementia. The results underline the importance of cardiovascular health, with emphasis on change over time, modulates cognitive aging, also in individuals with increased genetic risk.
The purpose of the fourth study was to develop an R-package for the method development that the project has generated. GcompBART implements a parametric version of Robin's g-formula based on Bayesian machine learning and can be used to estimate (causal) relationships between time-varying variables, e.g. change in cardiovascular risk or widowhood, on e.g. cognition or dementia, based on longitudinal data with time-varying confounders.
During the project, the PI (MJ) has collaborated with a number of national and international researchers, both biostatisticians and cognitive neuroscientists. In Studies 1 and 2, MJ collaborated with Professor Michael J. Daniels, at the University of Texas, Austin and in Study 2 also with Sara Pudas at Umeå University. MJD is an internationally recognized expert on methods for non-random dropout and his participation has significantly strengthened the method development in the project. During the project, Prof. Michael Daniels from the University of Florida, USA, has been on a 4-day exchange at CEDAR. In Study 3, MJ collaborated with several researchers at the Umeå Center for Functional Brain Imaging, Umeå University (UFBI), including Professor Lars Nyberg, PI for the Betula study and UFBI. In the fourth study, MJ developed software in the form of an R package.
The project's three main conclusions, which are described in detail in the articles related to the project (see separate publication list), are the following: i) Cognitive aging and dementia The results underline the importance of longitudinal data for studies of risk factors' effect on cognitive aging. Specifically, the results show that maintaining cardiovascular health over time reduces the risk of cognitive impairment and dementia, even for individuals with increased genetic risk. Our results also confirm previous studies, that memory in adulthood is largely unchanged until about 60 years of age and a sharp decrease in episodic memory is seen after that. ii) Population estimates By combining data from a longitudinal cohort study with Swedish register data on the target population, we have been able to show that we overestimate memory performance if we do not take into account differences in background factors between the sample and the population. iii) Method development Three advantages of the method development we have done are that i) no model selection is required, ii) can easily handle complex relationships and iii) the method can effectively handle selective dropout / inclusion and survival.
The research in this project has generated a large number of research questions that have resulted in new projects and project ideas. In collaboration with researchers at UFBI, we will in a new project investigate cardiovascular risk factors in relation to dementia based on brain imaging data. A further development of the project is that the project manager intends to continue to study dynamic connections between cardiovascular, social and behavioral risk factors and cognitive aging. An important contribution is that we will apply the methods we have developed in this project to answer the new questions.
Within the framework of the project, three articles have been published in highly ranked international journals with open access and peer review. All projects have also been presented at international conferences and at national seminar series. The method development has been implemented in an R package (GcompBART) which is freely available via https://github.com/m4ryjo/GcompBART.
In Josefsson and Daniels (2021), the purpose was first, to study the effect of becoming a widow / widower on episodic memory and second, to develop a better method for investigating this. We used longitudinal data from the Betula study, which allowed us to follow individuals over time on both widow status and memory. We included people who were married at baseline and then compared memory performance for individuals who became widows to those who were still married. The results did not show a significant difference between the groups. We used G-computation, which is a method for estimating a (causal) effect of a variable that varies over time (in our case widowhood status). Our method, unlike previous studies, can adjust for selective survival and non-ignorable dropout and implements Bayesian machine learning instead of standard regression models. The results underline the importance of longitudinal data for studying the effect of risk factors on cognitive aging, and that machine learning is an effective and simple tool for estimating these models.
In Josefsson et al. (2021) the aim was to estimate cognitive aging over the life span. The validity of previous longitudinal studies is unclear since i) the samples are often not representative of the population, ii) they did not take into account selective participation, and iii) did not control for practice effects. Here we combined cohort data for the sample with register data for both the sample and the target population to adjust for selective participation, and further extended the method in Study 1 to adjust for non-random dropout, death and practice effects. The results show that cognitive performance is overestimated with previous methods but the change over the life span is quite similar, i.e. we see that the memory is largely unchanged until about 60 years of age and a sharp decrease in episodic memory after that.
In Farnsworth von Cederwald et al. (2022) the aim was to investigate whether changes in cardiovascular risk profiles over time predict dementia and/or episodic memory impairment. As a measure of cardiovascular risk, a composite measure was used based on systolic blood pressure, diagnosis of hypertension, BMI, diabetes, smoking, age and gender. Three groups were identified, groups with stable, average and impaired cardiovascular health. Modeling was done with Bayesian machine learning for survival data with competing events (ie dementia and death). The results showed that individuals with impaired cardiovascular health were at significantly higher risk for both vascular dementia and Alzheimer's disease at older ages; but also an increased risk of episodic memory impairment already in middle age. For carriers of the risk gene Apoe E4, a maintained cardiovascular health can significantly reduce the risk of dementia. The results underline the importance of cardiovascular health, with emphasis on change over time, modulates cognitive aging, also in individuals with increased genetic risk.
The purpose of the fourth study was to develop an R-package for the method development that the project has generated. GcompBART implements a parametric version of Robin's g-formula based on Bayesian machine learning and can be used to estimate (causal) relationships between time-varying variables, e.g. change in cardiovascular risk or widowhood, on e.g. cognition or dementia, based on longitudinal data with time-varying confounders.
During the project, the PI (MJ) has collaborated with a number of national and international researchers, both biostatisticians and cognitive neuroscientists. In Studies 1 and 2, MJ collaborated with Professor Michael J. Daniels, at the University of Texas, Austin and in Study 2 also with Sara Pudas at Umeå University. MJD is an internationally recognized expert on methods for non-random dropout and his participation has significantly strengthened the method development in the project. During the project, Prof. Michael Daniels from the University of Florida, USA, has been on a 4-day exchange at CEDAR. In Study 3, MJ collaborated with several researchers at the Umeå Center for Functional Brain Imaging, Umeå University (UFBI), including Professor Lars Nyberg, PI for the Betula study and UFBI. In the fourth study, MJ developed software in the form of an R package.
The project's three main conclusions, which are described in detail in the articles related to the project (see separate publication list), are the following: i) Cognitive aging and dementia The results underline the importance of longitudinal data for studies of risk factors' effect on cognitive aging. Specifically, the results show that maintaining cardiovascular health over time reduces the risk of cognitive impairment and dementia, even for individuals with increased genetic risk. Our results also confirm previous studies, that memory in adulthood is largely unchanged until about 60 years of age and a sharp decrease in episodic memory is seen after that. ii) Population estimates By combining data from a longitudinal cohort study with Swedish register data on the target population, we have been able to show that we overestimate memory performance if we do not take into account differences in background factors between the sample and the population. iii) Method development Three advantages of the method development we have done are that i) no model selection is required, ii) can easily handle complex relationships and iii) the method can effectively handle selective dropout / inclusion and survival.
The research in this project has generated a large number of research questions that have resulted in new projects and project ideas. In collaboration with researchers at UFBI, we will in a new project investigate cardiovascular risk factors in relation to dementia based on brain imaging data. A further development of the project is that the project manager intends to continue to study dynamic connections between cardiovascular, social and behavioral risk factors and cognitive aging. An important contribution is that we will apply the methods we have developed in this project to answer the new questions.
Within the framework of the project, three articles have been published in highly ranked international journals with open access and peer review. All projects have also been presented at international conferences and at national seminar series. The method development has been implemented in an R package (GcompBART) which is freely available via https://github.com/m4ryjo/GcompBART.