08:00 - 09:30 Uhr
Melanchthonianum HS B
Vorsitz: Schilling, Oliver (Heidelberg); Zimprich, Daniel (Ulm)
Diskutant: Ziegelmann, Jochen Philipp (Berlin)
Aging research faces a number of methodological challenges. In the symposium, some of these methodological challenges and possible remedies will be presented and discussed. In a first presentation, Scott Hofer and colleagues focus on the importance of replication and cross-validation across heterogeneous longitudinal data from different sources. Conceptual and statistical models at the construct-level are evaluated by comparing alternative models of change, measurement harmonization and construct-level comparison, retest effects, distinguishing and contrasting between-person and within-person effects across different studies. Daniel Zimprich offers a local constant smoothing approach of latent variables. A possible application of such an approach is the investigation of different degrees of measurement invariance across age, which will be demonstrated using both simulated and real data. Oliver Schilling’s presentation touches another important issue in longitudinal aging research, namely, the modeling of highly skewed data using generalized linear mixed models (GLMMs). Different from typical linear mixed models with their assumption of normally distributed residuals, GLMMs offer a more flexible range of distributions from the exponential family. Employing the SAS procedure GLIMMIX, the usefulness of GLMMs for analyzing skewed longitudinal data will be demonstrated. Eventually, Tanja Kurtz in her presentation shows how genuinely non-linear models can be applied to model individual differences verbal learning data on the latent variable level using three parameters to describe performance changes: Initial level, learning rate, and asymptotic performance. In addition, she demonstrates how these parameters can then be linked to predictors of verbal learning in young and old adults, namely, processing speed and working memory. Jochen Ziegelmann discusses the presentations from both a method-oriented and applied perspective.
Scott M. Hofer (University of Victoria and Oregon Health & Science University), Andrea M. Piccinin (University of Victoria), Graciela Muniz-Terrera (University College London), & Philippe Rast (University of Victoria)
The analysis of longitudinal observational data can take many forms and requires many decisions, with research findings and conclusions often found to differ across independent longitudinal studies addressing the same question. Differences in measurements, sample composition (e.g., age, cohort, country/culture), and statistical models (e.g., change/time function, covariate set, centering, treatment of incomplete data) can affect the replicability of results. The central aim of the Integrative Analysis of Longitudinal Studies of Aging (IALSA) research network (NIH/NIA P01AG043362) is to optimize opportunities for replication and cross-validation across heterogeneous sources of longitudinal data by evaluating comparable conceptual and statistical models at the construct-level. I will provide an overview of the methodological challenges associated with comparative longitudinal research, including the comparability of alternative models of change, measurement harmonization and construct-level comparison, retest effects, distinguishing and contrasting between-person and within-person effects across studies, and evaluation of alternative models for change over time. These methodological challenges and recommended approaches will be discussed within the context of reproducible research.
Daniel Zimprich (University of Ulm)
Typically, if measurement invariance of a factor model is examined with respect to age using a multiple group model, an age-heterogeneous cross-sectional sample is split into a few distinct age groups. Such an approach has two drawbacks, however. First, information is lost by transforming the continuous age variable into a categorical variable. Second, the way the sample is split into age groups oftentimes appears arbitrary and driven by the goal of subsamples being large enough to warrant factor analysis while at the same time arriving at as many equal-sized samples as possible. As an alternative, local constant smoothing of factor models provides more fine-graded insights into possible changes of the parameters of interest (factor means, variances, and covariances) across age. Compared to traditional local constant smoothing approaches with their focus on manifest variables, the specialty of the approach to be presented lies in the fact that latent variables are subject to smoothing. To do so, the weighting of individual observations and the bandwidth for the age variable have to be transferred into factor models. The approach will be demonstrated using both simulated data with known properties and real data from research on a questionnaire tapping functions of autobiographical memory.
Oliver Schilling (University of Heidelberg)
In research on developmental processes, the mixed/multilevel model approach to the analysis of intra-individual change has become widespread, including the use of “conventional” mixed model estimation algorithms that rest upon the assumption that the response is normally distributed. However, research on aging is often concerned with outcomes that typically show highly skewed distributions – such as, e.g., measures of depressive symptoms and negative affect – that may not meet the normality assumption. Less known among researchers in this field, so-called generalized linear mixed models (GLMM) comprise algorithms for mixed modeling of data that can have any distribution in the exponential family, including the beta, gamma, and other distributions that may better fit with skewed response variables. The presentation is aimed to demonstrate the GLMM approach to model intra-individual change in highly skewed outcome data, using SAS procedure GLIMMIX. Scores of depressive symptoms and negative affect from several longitudinal aging studies will be analyzed for demonstration, and results based on normal, gamma and beta distribution of the response variable will be compared.
Tanja Kurtz (University of Ulm)
Previous studies show that persons in old age differ regarding recall performance of verbal material. Moreover, previous studies show that basal cognitive abilities such as processing speed and working memory explain, at least to a substantial degree, individual differences in verbal learning in old age. In order to investigate these relations and possible underlying mechanisms more precisely it seems necessary to analyze the role of basic cognitive abilities for verbal learning in younger and in older age. Such analytical approach then addresses the issue of dedifferentiation---do relations between basic cognitive variables and verbal learning increase in older age? Aiming at analyzing mechanisms of verbal learning by using this approach, the present study analyzes whether basic cognitive abilities are related with individual differences in learning in younger and older age. Therefore, a verbal learning task including five trials is implemented for a younger (N = 205; 18-30 years of age) and an older age-group (N = 364; 65-80 years of age). Within both age-groups, individual differences are modeled via latent growth curve models with three parameters: Initial level, learning rate, and asymptotic performance. These parameters are then related with the individual processing speed and working memory capacity. Overall, results show that the relation between basic cognitive abilities and verbal learning is more pronounced in younger age. This result indicates that the mechanisms of verbal learning may differ for younger and older persons---by taking into account that individuals differ regarding their verbal learning performance within their age-group. Results might further be discussed from a dedifferentiation perspective: Why are basic cognitive abilities related to a higher degree with verbal learning in older than in younger age?