The upcoming Conference on Innovations in Trauma Research Methods (CITRM; November 3-4, 2006, West Hollywood, Calif.), will feature a session on new methods for analyzing intensively measured time series data. Intensive data consist of frequent assessments made over many short time intervals within participants. As a prelude to that session, Susan Doron-LaMarca, PhD (National Center for PTSD and Boston University School of Medicine), offers this brief introduction to analysis of chronic fluctuation using intensive longitudinal data.
To examine change in posttraumatic symptoms over time, researchers are increasingly using longitudinal designs that typically include repeated assessments of an individual’s mental health status. This produces what is called ‘nested’ or ‘multilevel’ data in which multiple units or observations at a ‘lower’ level are grouped within ‘higher’ or ‘upper’ levels. A classic example of nesting from educational research is children nested within classrooms nested within schools. A corresponding example of nesting using a trauma-related longitudinal design could include, for instance, repeated assessments of PTSD over time among survivors of motor vehicle accidents. Here, the repeated observations of PTSD symptoms (and perhaps the times at which they are assessed) are nested within each study participant, representing lower level data, and enduring characteristics of each participant that need only be assessed once (e.g., gender or age at time of the accident) represent higher level data.
In turn, participants could be nested within some other grouping factor (e.g., therapy group) to add yet another level. Using this type of nested or multilevel data, researchers can ask questions about (a) overall increase or decrease in PTSD symptoms over time (e.g., slope) within persons; (b) whether the change in PTSD represented by the slope for each person is constant (e.g., linear) or changing (e.g., curvilinear); and (c) whether there are differences in change in PTSD (e.g., slope) across persons as a function of the higher level characteristics, such as gender or age, or further differences in change as a function of therapy group.
Traditional methods may estimate group-level change by comparing, for example, the mean PTSD score across individuals at Time 1 to the mean PTSD score across individuals at Time 2. This can result in means across individuals at each time point that are nearly equal, essentially reflecting a lack of overall change. Yet, at the individual level, there may be evidence of substantial change (e.g., PTSD scores are increasing for some participants, while decreasing for others) that is better captured by newer methods that estimate an individual trajectory of change in PTSD.
New longitudinal data collection techniques allow participants to make frequent and repeated self-reports about their symptoms and behaviors as they occur. These new modes of data collection include telehealth and electronic diary methods in which participants enter self-reports into hand-held computers (PDAs) as they go about their daily activities. For example, following a motor vehicle accident, participants could enter self-reports of PTSD symptoms into PDAs multiple times per day. The resulting data structure would consist of many observations (e.g., 30 to 100 or more) assessed over short time intervals for each participant. These data are referred to as time series data or intensive longitudinal data.
Intensive longitudinal design and the resulting data go hand-in-hand with novel multi-level statistical methods that allow the examination of prospective relationships among variables in new ways, thus expanding the range of questions that researchers may address regarding change. In addition to individual growth trends, researchers may want to examine a different type of posttraumatic process: chronic fluctuation, consisting of increases and decreases over time without a discernible trend toward growth or decline. Research questions might address the weekly course of symptoms in a chronic disorder, daily fluctuation in mood, or hourly assessment of behaviors such as smoking. Capturing aspects of chronic fluctuation can be challenging, since it lacks an overall trend. In contrast to the growth curve modeling example presented above that examined change in PTSD as a function of the repeated time measure, one might wish to explore whether a person’s standing on an outcome of interest can be predicted from his/her standing on the same variable at a prior occasion. One can determine, for instance, the extent to which PTSD symptoms assessed at the previous occasion predict PTSD for each person at the next occasion.
An important difference in this approach from other longitudinal methods is the focus on the association between each person’s prior status and his/her later status: Each person’s has his/her own regression coefficient that estimates the strength of this relationship (i.e., the association between prior and later PTSD). This coefficient is a ‘score’ or variable in and of itself, an individual differences characteristic worthy of study and prediction. This makes it possible, in a multilevel framework, to ask whether differences between individuals on risk or resilience factors (e.g., the aforementioned enduring characteristics, such as gender or age at time of trauma) influence or moderate these within-person associations. Of course, there are elaborations on the methods that permit more complex and interesting research questions, perhaps related to the dynamics among different PTSD symptoms over time within persons, or fluctuation as a function of the influence of other concomitantly measured variables, such as ongoing life stressors or social support.
The method described above and other methods for intensive longitudinal data are presented in a recent book, "Models for Intensive Longitudinal Data," edited by Theodore Walls and Joseph Schafer and published by Oxford University Press in 2006.