Overnight prediction issue

[Chang Liu]

Hi everyone,

I'm writing to inquire about the usage of the Gimme package in R, specifically regarding the prevention of overnight predictions when analyzing EMA data collected at multiple time points per day (e.g., four times a day with a 3-hour interval) over a period of time using Gimme. My understanding is that, in this context, one measurement (the last measurement during the day) should not predict the subsequent measurement overnight (i.e., the first measurement on the following day) due to different time intervals.

As most of the Gimme tutorials I've come across focus on data collected at a daily interval, which may not encounter this issue (the need to prevent overnight prediction). I would greatly appreciate it if you could share any relevant resources and code that address the issue of overnight prediction when using Gimme or demonstrate the application of Gimme to EMA data collected at multiple time points per day over an extended period. 

Thank you for your support.

Cheers,

Chang

[Katie Gates]

This is a great question and an increasingly common issue. 

There are two approaches that I know of. 

1. Interpolate the data so that the observations are equally spaced. Aaron Fisher does that with his data. Here is a paper where he provides code and the data: https://osf.io/5ybxt/. 

2. Put "NA" in between the final evening observation and the next morning observation. This will prevent this lag from being included when arriving at cross-lag and autoregressive estimates. So then all lagged relations will be within day. The contemporaneous relations will be unchanged. I believe this is what was done in Assessing Temporal Emotion Dynamics Using Networks by Bringmann et al., 2016. 

Each approach has pros and cons. I'm curious what others think is the best option. 

hope this helps, 

Katie

[Chang Liu]

Dear Katie,

Thank you for your response; it was very informative. I do have some additional inquiries:

Incorporating Sleep-Related Measures into the Gimme Model: One of my concerns is the inclusion of sleep-related measures in the Gimme model. These measures are often taken at different time scales compared to other items, with sleep having a one-day time window and other items (e.g. affect) having a three-hour time window. Additionally, we may have fewer observations for sleep measures. I’m wondering if there are any recommended practices/existing examples that include Sleep-Related Measures when using Gimme.

Quantifying Node Centrality in the Gimme Network: I would like to know how to quantify the centrality of nodes in the Gimme network. Is there an automated process or code available for this, similar to the centralityTable/centralityPlot function in qgraph?

Conducting Comparison Tests for the Gimme Network: I'm also interested in how to conduct comparison tests for Gimme networks. Specifically, I'd like to compare the network structure invariance, global strength invariance, edge invariance, and centrality. 

Resources and code for these questions would be greatly appreciated.

Cheers,

Chang

[Katie Gates]

Good questions. I don't have all the answers but can point you in some directions. 

When observations are taken at unequal frequencies, people sometimes use the same value for multiple time points. For instance, the hours of sleep the night before will likely impact the person throughout the entire day. So perhaps the same value can be used for the other measures which are taken every 3 hours during the day. Another option is a weighted mean, where the observation in the AM are given the sleep value from the prior night, and the subsequent observations are a combination of the prior night and the next night. This might not make much sense for a sleep variable, so I'd recommend the first option I suggest. 

For quantifying node centrality, I would recommend reading Laura Bringmann's work on this. I've included a couple of citations below. In short - igraph can be used on gimme output to obtain centrality measures. The Bingmann et al. 2016 paper describes how one can do that when both lagged and contemporaneous matrices are obtained. 

For conducting comparison tests, yes you can compare people based on the measures mentioned (e.g., centrality). Those measures can be obtained using igraph. For testing structural and edge invariance, I would point you to the literature on testing for configural measurement invariance. The same concepts apply here. I haven't done work on this so can't expand on how to do it beyond that. 

Best,

Katie

Bringmann, L. F., Pe, M. L., Vissers, N., Ceulemans, E., Borsboom, D., Vanpaemel, W., ... & Kuppens, P. (2016). Assessing temporal emotion dynamics using networks. Assessment, 23(4), 425-435.

Bringmann, L. F., Elmer, T., Epskamp, S., Krause, R. W., Schoch, D., Wichers, M., ... & Snippe, E. (2019). What do centrality measures measure in psychological networks?. Journal of abnormal psychology, 128(8), 892.