On this sunny summer Monday morning, we should talk a little bit about estimating and interpreting intensive longitudinal data models.

I'm mostly thinking ambulatory assessment (EMA, ESM, daily diary, etc), but the point I want to make is general.

To lag or not to lag? 1/16

Intensive longitudinal designs involve sampling repeatedly on some schedule (e.g., every two hours, daily, 6x randomly per day).

Collecting data this way gives you (usually multivariate) time series, which can be analyzed in a number of ways.

2/16

Ambulatory assessment buys you (at least) 3 things:

1. Avoid retrospection
2. Ecological validity (it's what happens in the real world, not the artificial lab)
3. It allows you to (potentially) model a process as it unfolds

This last feature is the focus of this thread. 3/16

When estimating models using the time series data, one key distinction is whether associations among variables are "contemporaneous" (ie within same assessments) or "lagged" (a variable at one assessment predicting variable at next assessment).

4/16

There is a tendency in the field to prize lagged associations.

I get it. Being able to forecast or predict what comes next is basically the holy grail of the behavioral sciences.

But lagged effects are NOT inherently more valuable.

5/16

Too often I hear contemporaneous associations written off as "they are just cross-sectional."

This ignores at least 3 things:

1. They are dynamic, emerging from fluctuations over timepoints.
2. The true timing of the process of interest.
3. The wording of the variables.

6/16

Let's look at these.

1. in ILD, contemporaneous still means "dynamic!" The only way you can get these associations is by the fluctuation of variables from occasion to occasion of measurement.

That is qualitatively different than cross-sectional in the traditional sense.

7/16

1. (cont) This can lead to big differences in patterns, such as affect structure within- and between-person, which is often aligned towards valence vs. orthogonal PA vs. NA

At any moment rare to feel both + and -, but people who feel more + are not necessarily less -.

8/16

1. (cont) Or in population you find that neuroticism and conscientiousness are negatively correlated, but in the moment, managers who are behaving more conscientiously experience more negative emotions. Let's get those problems solved!

9/16

1. (cont) Or people who spend more time seeking for meaning in life experience are less happy and satisfied, but when someone seeks out meaning they experience more happiness and satisfaction in the moment.

So, these dynamic associations are different in important ways.

10/16

2. Timing matters. We basically have little idea how long it takes for processes to unfold for very complex behaviors (e.g., negative emotions and suicidal thoughts).

If you don't know the timing of a process, it seems very strange to prioritize a lagged association.

11/16

2. (cont) Why? Well, because lagged associations capture any process that is LONGER than your assessment schedule, but contemporaneous capture any process that is SLOWER than your assessment schedule. Granger told us this in 1969.

12/16

2. (cont) So, unless you think your process is slower than your assessment interval, there's no strong reason to argue for lagged values.

e.g., Do we think it takes >2hrs on average to go from negative emotions to self-regulating behaviors? Or is it quicker?

13/16

2. (cont) That's the question you have to ask yourself. And, importantly, reflexively assuming only lagged effects tell you about a process is to be avoided.

Unfortunately, we're often in the dark, so may have to explore both.

14/16

3. Last point, and its minor, but you need to pay attention to item wording. If affect is "how are you feeling now?" but some behavior (impulsivity) is "since the last prompt", you probably want to lag if you care about affect->impulsivity. Otherwise you may be backwards.

15/16

OK, if you've made it this far into this mornings micro-lecture/rant, thanks for listening.

Good luck collecting your ILD and modeling it!

16/16