Thread by @drrodebaugh

This thread is a live-tweet of Novel Analytic Approaches over Time (chair, Michael Halquist) at #aps19dc. I'm live-tweeting this one while @mcmullarkey is live-tweeting one on Recent Advances (chairs @EikoFried and @talyarkoni) bc several people want to see both!

Revelation: Danny Pine is not here, that was a misprint.

Clarissa Low is up first, talking about smartphone and wearable sensor data. She first reviews how smartphones and sensors have become common, nearly ubiquitous, across ages, ethnic groups, etc.

Smart phones can passively capture data, allow personalized interventions [my silent question: but how do we know how to best personalize?]. Dr. Low has been using the Aware framework, which is new to me--looks pretty cool: awareframework.com

This framework can grab data from loads of sensors (more than shown in this picture). It also has a plug-in to detect whether a conversation is happening, which seems super useful. Other frameworks are also available.

Cornet and Holden, 2018, reviewed studies on passive sensing for health and well-being, sounds worth a look if you're interested. Dr. Low shares a sensing of quality of life study for patients undergoing chemotherapy for GI cancer. Sensing by smartphone and fitbit.

Data obtained, they first extract features based on a 24 hour time window. (So much data, have to pick some way to extract data.) Much of this is over my head, but I appreciate her discussion of missing data--it sounds like missing data could use more work in machine learning.

They used supervised learning to try to create a model to detect quality of life (e.g., high versus low symptom days, based on self-report of 12 symptoms). Were able to predict well both in groups and in individuals, although it sounds like those needed different models.

I'm not listing all of the sensors that ended up being important: More useful is to know it was a lot, it was not all obvious, and it seemed to vary by type of group and also by individual.

Her lab is working on lots of possible applications, including loneliness, which is particularly exciting to me. Dr. Low is on twitter, so I'm belatedly tagging her here and correcting her name: @carissa_low Apologies for the incorrect first name the first time!

Next Dr. Halquist on state-space models. These models come from engineering with the intent to track spacecraft over time. I did not know this and am glad I understand why the models are called that!

In the model, you have latent states (following a Markov process) and these unfold over time. You also have observed measurements, but once you account for the latent states, these measurements are otherwise unrelated. Somehow, the latent states relate to observations (math)

The math includes an "evolution parameter" that specifies how the latent state is changing over time. For example, affect could be pulled back toward an attractor, but also have random, noisy aspects. There are multiple types of evolution functions.

Also in the math, an "observation function," which is how the states relate to observable data. This can be simple or complicated.

Exciting for me: these models can describe nonlinear dynamics, as well as the familiar linear ones. They also push us to write out formal equations for what we think is happening (@EikoFried will like that one).

Dr. Hallquist provides examples from personality pathology. First, physiological coregulation (linkage between two people, as perhaps in attachment relationships, in which people pull each other toward a shared baseline). Personality pathology gets in the way of this.

Example 2: emotion dysreg in borderline PD. This is Hallquist and Dombrovski (2019) and there are too many graphics and math for me to do it justice. Short version: When disrupted by fearful faces, people with borderline PD tend to make poorer decisions.

Ends on a "shameless" plug for book on Research Methods in Clinical Psychology, which looks great (co-edited with @aidangcw) and includes lots on longitudinal data.

Now @DejonckheerEgon on emotion dynamics. I'm going to slow down on tweeting because, actually, my hands hurt a little. #NotARobot

Fortunately @DejonckheerEgon has already put on the web related thoughts here: socialsciences.nature.com/users/210791-e…

In short, we have lots of proposed measures of emotional dynamics (how emotions change over time, relate to themselves, relate to other emotions). Are all/any of them really different? Do any of them beat mean levels of affect in predicting well being?

Questions tested in 15 studies (1777 participants!). Merin Mestdagh presents methods and results. All the dynamic measures are related, some more than others. In particular, variation measures form a clump. They also don't tend to beat the mean as predictors.

*Some* evidence for dynamic measures beating the mean in predicting Borderline PD symptoms. Not a lot, but some. But it turns out that most of that is attributable to the standard deviation--controlling for that, the dynamics don't predict anything.

Now Matthew Barstead on momentary emotional experience and psychopathology. Ooh, look, multilevel variance decomposition. Focus: dispositional negativity (more or less, neuroticism or trait negative affect). It's always neuroticism.

Argues that we need to know more about how neuroticism affects reactions to daily stressors before we'll have a good idea how to intervene. [My silent question: Don't we already have interventions that we know work? What about those?]

Rights and Sterba 2018 (psych methods) is where to look for the methods used here. Didn't find neuroticism to predict reactivity (a surprise) but did find more exposures to stressful events and more general negative affect.

[Missed some results while typing, rookie error] Neuroticism makes people generally feel worse, also have more troubles, but doesn't necessarily lead to greater reactivity to events per se. You see the same pattern, but with some evidence for reactivity, in higher N folks, tho

Call for more of an examination of positive mood and positive events, which remain understudied in this area [agree!]. End of talks!

Question for @DejonckheerEgon (from Sherryl Goodman): was all of this because of the populations studied? Age? Clinical groups absent? Actually, no, there was a wide age range and lots of clinical participants.

Question: Is this time series symposium telling us everything is all about the mean? @carissa_low reports that in her studies it's not always the mean that predicts, sometimes it's change. Part of it may be that what we're trying to predict is a trait-like symptom measure.

[I say a thing]--discussion moves to psychometrics and that we need to find ways to avoid things like floor and ceiling effects in EMA measures.