Prof. R.R. Meijer & Prof. M.E. Timmerman
On September 22nd, 2016, Tanja Krone will defend her thesis entitled
Some Notes on Bayesian Time Series Analysis in Psychology
(Thesis and summary not available yet)
Understanding human behavioural processes with Bayesian dynamic models
The use of research designs with intensive measurements across time for individual subjects is becoming increasingly popular in psychological research. Such designs are necessary to achieve insight into the extremely complex phenomena of human behaviour like emotions (Scherer, 2009) and psychopathology (Frank et al., 2005). This complexity finds expression in behaviour fluctuating across time. Since those fluctuations depend on contextual and interindividual differences, understanding the underlying dynamics is extremely challenging. With this challenge, statistical time series analysis can be of great help. In general, the analysis of time series data serves either or both of the two main purposes:
(i) to study the time series itself to gain insight into the processes underlying the data;
(ii) to forecast, that is, to use observed data to predict unobserved future data.
When studying the time series, random noise is separated from systematic patterns in the data (e.g., Box et al., 1994). The systematic component is usually modelled, for example, by splitting into seasonal and trend components. This is relevant, for example, to identify whether a patient suffering from winter depression shows less symptoms of depression after a therapy, apart from the usual seasonal fluctuations. The main goal of forecasting models is to predict unobserved outcomes on the basis of observed history. Examples include statements on the density of traffic and on the necessary time for a patient to receive treatment before successful recovery.
Although the merits of the principles underlying time series analysis have been shown convincingly in psychology (e.g., Lodewyckx et al., 2011), the models used so far suffer from important limitations. As will be discussed below, the number of dependent variables and their nature to include in the analysis is limited. Furthermore, the models are static, rather than dynamic in nature. Those limitations imply that important dynamics will be kept hidden. Resolving those limitations would be extremely helpful, since understanding the dynamics offers a key to influencing, which is of utmost importance in diagnosis and planning psychological interventions. Furthermore, forecasting can be very useful, for example in forensic psychiatry (e.g., to predict aberrant behaviour), or in youth care (e.g., early tracing of anomalies in development). To resolve the limitations of the time series models used so far, we will extend the linear multiregression dynamic model (LMDM; Queen et al., 1993, 2007, 2008, 2009) to more general Bayesian dynamic models (BDMs). The LMDM, which has been successfully applied to traffic forecasting, has a number of favourable properties that make the model eminently suitable for psychological time series. We will develop some necessary theoretical extensions, and apply the variant developed to empirical examples from typical psychological time series research. To examine the value of the BDMs in relationship to currently popular time series models, we will perform a comparative study based on simulated and empirical data.
NWO(Netherlands Organisation for Scientific Research)