Developing an information-theoretic predictive factor analysis method with application to transdiagnostic psychometric and neurocognitive data

Statistics & applied probability
Factor analysis/information theory/dimensionality reduction/psychometrics/psychology

Large scale data sets are increasingly available within the life sciences & medicine, in which many diverse predictive variables are obtained & related to a health-related outcome.Information theory provides a way to directly quantify representational interactions between pairs of predictors.These interactions can take the form of redundancy or synergy.Redundancy quantifies overlapping predictive power within individuals.Synergy indicates that the relationship between 2 predictors within an individual is itself informative about the outcome;you obtain better predictive performance when considering both predictors together than you would if you combined the prediction made from each.This project aims to develop a new information-theoretic factor analysis. Applying recently developed information theoretic tools in a pairwise fashion will allow us to build optimum sets of predictive features by combining synergistic predictors which can be grouped into common predictive factors by clustering redundant features.A cluster of redundant features represents a set of predictors which provide the same predictive information about the outcome across individuals & therefore represents a predictive factor.This methodology will be developed with a large new data set,collected online,to investigate transdiagnostic psychiatric symptom dimensions & their relationship to different aspects of cognition.Traditionally, mental illness has been classified & diagnosed according to formal taxonomic systems,whereby sets of psychopathological symptoms are organized into discrete conditions that correspond to specific psychiatric diagnoses.However,there is increasing evidence that the symptom space is not discrete but dimensional.Importantly,the assumption of the existence of the currently accepted discrete diagnostic categories may be hindering the identification of underlying neural & cognitive substrates & causal mechanisms of symptoms associated with psychopathology.On account of these limitations & concerns regarding the validity & utility of the categorical approach,psychiatric researchers are moving toward a transdiagnostic approach to defining mental health problems.Their aim is to map variations in psychopathology in the general population,to derive valid & psychometrically sound dimensions of mental health & subsequently,aid identification of underlying causes & mechanisms.Heterogeneous application,however,has resulted in a lack of replication and over-enthusiastic interpretation of results & has not yet provided a superior alternative to the current discrete framework. In order to provide the symptom dimensions approach with a robust statistical foundation,this project will collect a large data set to replicate existing results based on factor analysis of psychiatric symptom questionnaires.This will provide a benchmark from a topical & promising research area to which to compare the novel information theoretic approach.We will employ a comprehensive battery of behavioural tasks & psychometric questionnaires across a large online general population sample,to investigate the relationship between transdiagnostic symptom dimensions & neurocognitive task performance.Further, we will apply traditional factor analysis & our novel information theoretic factor analysis to our battery of cognitive tasks to determine if there are common behavioural factors & if and how these relate to the transdiagnostic symptom dimensions.This project therefore has two main aims. 1st-to produce a valuable new data set to investigate transdiagnostic symptom dimensions & their relation to different aspects of cognition, as well as looking for cognitive factors across a range of tasks in a large general population sample. 2nd-to develop a novel information-theoretic factor analysis & compare this to traditional factor analysis in this application.