Many researchers know that understanding how to use / analyze your data is invaluable when conceptualizing clear, specific research questions and hypotheses. The goal of this post is to enhance readers’ understanding of statistical modeling strategies for the types of data that NeuroUX can help you collect, so you can ask the best questions possible for your study!

EMA and mobile cognitive testing data are commonly collected in measurement bursts. For example, a study using the NeuroUX platform may involve a burst of 7 consecutive days of daily paired EMA surveys and mobile cognitive testing sessions, and this burst may repeat every 6 months. The data that results from this study are longitudinal and they are nested. Their longitudinal nature is intuitive – i.e., the study collected repeated measures (EMA survey responses & mobile cognitive test scores) over time. The term nested describes the structure of the data whereby EMA/mobile testing sessions from each burst (Session1 to Session7) are nested within participants (ID1 to IDn) (Figure 1).

Figure 1. Nested longitudinal data, with individual EMA/testing sessions nested within participants (clusters)

These data allow us to examine two primary levels of analysis: (1) between-person relationships and (2) within-person relationships. Between-person relationships test interindividual variation, i.e., differences between people. These are the types of relationships that we examine with cross-sectional studies, including associations between in-lab cognitive performance and sex/gender, for example. Within-person relationships test intraindividual variation, i.e., how a person differs from themselves at different points in time. Longitudinal data, as described above, have the power to answer questions about within-person associations by examining variables that are expected to change over time (e.g., mood, attention) and are measured more than once.

There are several ways we might hypothesize certain variables to change over time within persons in a longitudinal study, including (1) consistent change in one direction or (2) fluctuation. An example of consistent change might include cognitive decline in older adults with dementia due to Alzheimer’s disease, whereby their performance on cognitive tests are likely to show consistent declines over time (Figure 2A). In contrast, an adult with ADHD may be more likely to show fluctuations in their cognitive performance / attention over time (Figure 2B). While you can examine hypotheses around either scenario, data collected from EMA + mobile cognitive testing within a single short burst is uniquely well-designed for questions about acute fluctuation.

Figure 2. Visual examples of different types of change over time

Using NeuroUX data from the example study design described above, you can test both between-person and within-person hypotheses, e.g.:

  • Between-person hypothesis: Do participants who have higher average EMA sadness ratings perform worse on mobile tests of working memory on average compared to participants with lower average sadness ratings?
  • Within-person hypothesis: When someone reports being sadder than usual, is their performance on working memory tests worse than usual? [Note: this examines a hypothesis about fluctuation!]

, where sadness ratings are a primary predictor and working memory performance is the outcome.

So how do we analyze these data to test those hypotheses? When we have nested (AKA hierarchical or clustered) data of any kind, they violate the assumption of independence for traditional regression models – that is, observations are not independent of each other when they are repeatedly taken from the same participants over time. Instead, we can use mixed-effects models (AKA multilevel models or hierarchical models), which model both fixed and random effects simultaneously and are necessary for distinguishing between-person from within-person relationships in your longitudinal data. [Note: While this is certainly not the only method to analyze longitudinal data, it is a common technique that is easy to implement and worth knowing about!]

For more details on mixed-effects modeling and recommendations on which R packages to use, reach out to our team at contact@getneuroux.com!

-----

This article was written by Dr. Emily Paolillo, an Assistant Professor at the University of California, San Francisco and edited by our co-founder, Dr. Raeanne Moore.