Thread by @NC_Jacobson

**Another #tweetorial**

Today, I had another presentation at @SAA2020aus on the importance of asking WHEN examining dynamic in intensive longitudinal data.

This is a summary for those who couldn't make it.

⚠️Warning⚠️
Again, I'm arguing against common practices.

(1)

Intensive longitudinal data allow great opportunity to examine relationships between varying states.

BUT, we have some practices that in this field which are frankly bizarre unscientific heuristics that inform both design & analysis.

Timing is often ENTIRELY overlooked

(2)

In designing active assessment studies when we don't know when an event happens we can use:

Signal-contingent: pinging people & asking to respond usually with either random or windowed random windows

Interval contingent: complete assessments at regular intervals in time

(3)

With dynamic processes, timing matters a LOT

If we mis-time our data collection, it has huge implications

Example: Movement of helicopter blades

Timing of assessments (i.e. shutter speed here) can make it look like they move very slowly, reverse direction of the blades

(4)

How do you choose when to sample?

Theory!
If you have a strongly suggestive theory about timing, great, this almost never happens...so what should you do?

⛔️Common practice!: Researcher X, Y, & Z each used that interval before (but they had no rationale)!
Do not do this!

(5)

Best practice!

Try to sample your process systematically – it’s much better to oversample than undersample

Incorrect timing will not allow us to test our theories

(6)

Time for the analysis

Common practice: Look at dynamic processes concurrently or one unit in time (i.e. lag) later

✅ Great, IF you have a strong theory about timing (but you probably don't)

❌Not okay in the vast majority of intensive longitudinal data work

(7)

Best practice: Investigate the times in which processes optimally predict one another

BTW, no you can't just eyeball higher-order lag dynamics (you can for lower-order dynamics), see the difference?

(8)

Options:

⛔Fit higher-order lags manually, false positives reign supreme
❓ Continuous time structural equation models, not bad but, not very flexible w/o higher order stochastic differential equations (which ridiculously hard to interpret)
❓Fractals, but hard to understand
(9)

Better alternative:

The differential time varying effect model

Fits smooth curves in exploratory way then uses gold-standard approaches to confirmatory modeling

Works with: 1 to many variables, 1 to many people, high missingness

Well-validated:
link.springer.com/article/10.375…

(10)

Better yet, it's highly automated

ONLY ONE line of code:
out=LAG("X",differntialtimevaryingpredictors=c("X"),outcome=c("X"),data=exampledat1,ID="ID",Time="Time",k=9,standardized=FALSE,predictionstart = 1,predictionsend = 10,predictionsinterval = 1)

(11)

The output is also pretty intuitive and pretty (at least in my mind)

(12)

Researchers: ACT NOW!

DESIGNING STUDIES:
Do: test theories w/ timing
Do: aim to oversample rather than undersample

ANALYZING STUDIES:
Do: Test lags if theorized (usually not)
Do: Explore higher order lags

(13)

You can find the DTVEM package here:
nicholasjacobson.com/project/dtvem/

And a simple tutorial here:
nicholasjacobson.com/post/illustrat…

Please contact me if you’re trying this out and have any questions!

(14)

All my slides are posted here:
nicholasjacobson.com/files/talks/SA…

Thanks to all of you suffering through another twitter tirade

(/end rant)