ADNI data is made available to researchers around the world. As such, there are many active research projects accessing and applying the shared ADNI data. To further encourage Alzheimer’s disease research collaboration, and to help prevent duplicate efforts, the list below shows the specific research focus of the active ADNI investigations. This information is requested annually as a requirement for data access.
|Principal Investigator's Name:||Rochelle E Tractenberg|
|Institution:||Georgetown University School of Medicine|
|Country:||United States of America|
|Proposed Analysis:||ADNI data analysis proposal. Submitted 7 January 2008. Rochelle Tractenberg, Georgetown University This proposal represents a two-part demonstration of the potential that latent variable modeling techniques can have in the biomedical domain. Specifically, the PI is a statistician and cognitive scientist with experience analyzing data from Alzheimer?s disease patients, who is developing expertise in latent variable modeling. The proposed analyses are intended to demonstrate how latent variable methods, which are typically used in social science applications and/or with variables that represent ?truly unobservable? constructs, such as ?intelligence? or ?depression?, can be useful with biomedical data. In his 1947 book, LL Thurstone demonstrated the method of factor analysis by obtaining length, width, and height (x, y, z) measurements on 20 randomly selected boxes. The measurements were permuted in twenty different ways (pp. 141-142), listed below. x^2, y^2, z^2, xy, xz, yz, (x^2+y^2)^1/2, (x^2+z^2)^1/2, (y^2+z^2)^1/2, (2x+2y), (2x+2z), (2y+2z), log x, log y, log z, xyz, (x^2+y^2+z^2)^1/2, e^x, e^y, e^z. These manipulations result in an ?augmented? data set that does not include the original three variables, x, y, and z. In 1975 Kaiser & Horst contributed to this problem by adding a 20x20 matrix of random measurement error, which had the effect of rendering the original 20x20 matrix nonsingular. Thus, the 20 manipulated variables, with added random error, are a realistic representation of a collection of variables that are interrelated, not precisely measured, and which represent a far smaller set of ?main? or common dimensions. By analyzing the box data in this early simulation, Thurstone demonstrated that factor analysis can recover ?true? dimensions, in his case, by showing that the manipulated variables loaded, as expected, on the dimension(s) from which they were obtained. For the first part of the proposed analysis, four imaging variables from each ADNI participant will be obtained, and manipulated in similar ways to how Thurstone originally did. Four variables are sought because the second part of the analysis is a demonstration of a latent variable model that permits simultaneous analysis of the four observed variables. The purpose of the analysis in the first part is to demonstrate that ?hard outcomes? such as white matter density, grey matter density, hippocampal volume, and overall brain volume, even if permuted and measured with error (as the error matrix will suggest), can be recovered by a latent variable analysis method. Unlike Thurstone and Kaiser & Horst, the latent variable analysis will be accomplished with freeware TETRAD (http://www.phil.cmu.edu/projects/tetrad/), which is a correlation constraint analysis instead of simply the dimensional reduction analysis that exploratory factor analysis represents. A secondary purpose of the first part of the analysis is to conduct a formal TETRAD analysis that will document how and why it works; many representations of TETRAD analyses in the literature are incomplete, inconclusive, or incorrect. The second part of the proposed analysis is to compare the statistical analyses that can be done with a set of four observed variables. The typical biostatistical analysis involves selecting the single best indicator as the dependent variable, and using multiple regression to examine the relationships between it and independent variables of interest, with or without covariate and interaction terms. A less focused approach would be to conduct a MANOVA, which permits all four indicators to serve as dependent variables, but which can require extensive post hoc analyses in order to determine what is different from what, and in which groups. Conversely, a latent variable model with four indicators representing different, measurable aspects of one latent variable, such as ?neurodegeneration? or ?disease burden?, will simultaneously analyze all four indicators (or however many are involved in the analysis) as representatives of a single underlying or common factor. In studies of Alzheimer?s disease, there are often many different neuropsychological test scores (?soft? outcomes) obtained, all of which are known or believed to represent ?cognitive function?, ?dementia?, or ?cognitive state?. Similarly, the ADNI study has collected a multitude of ?hard? outcomes, with the goal of determining which, if any, are reliable and valid indicators of some unobservable ? latent ? construct, process, or state. These outcomes themselves may be measured with error, but they certainly do not correspond wholly to any specific clinical outcome, a latent variable model involving multiple observed variables as indicators of the underlying variable of interest would be helpful. The second part of the analysis does not seek to find the ?best? model, only to use EQS 6.1 (Multivariate Software, Inc., 2005; Bentler and Wu, 1995) to demonstrate how latent variable methods can be used to construct reasonable models of underlying processes or states in a biomedical context. It would be ideal to obtain the data from both normal and MCI patients, since the latent variable method should be replicable across independent cohorts. The PI seeks both the data and also some assistance in selecting the best (most well-supported in the literature) four indicators that can be considered representative of a single underlying construct. However, Georgetown University has an imaging center and some assistance can be obtained there. The analyses could be used to create a poster and a presentation for an upcoming AD conference, but the main goal is a manuscript before the end of 2008. References Bentler PM and Wu E. (1995). EQS structural equations program manual. Encino, CA: Multivariate Software Inc. Kaiser HF and Horst P. (1975). A score matrix for Thurstone?s box problem. Multivariate Behavioral Research, January:17-25. Thurstone LL. (1947). Multiple Factor Analysis. Chicago: University of Chicago Press.|