+ This work was funded by the NIH. This shows how agencies don't contribute to public health but also results in new tools and methods for data collection and AI development.

Congratulations to all the co-authors, @shirlen3.bsky.social , Jixin Li, @genevievedunton.bsky.social , Wei-Lin Wang, Don Hedeker, and Stephen Intille. The details of the TIME study can be found here: timestudydocumentation.github.io/docs/build/h... (Reach out for questions or more info 😀)

This is one of the most intense EMA data collection studies leveraging smartphones and smartwatches. The μEMA method can be leveraged reliably for large-scale personalized data collection, just-in-time adaptive interventions, and human-in-the-loop ML that requires human feedback.

Finally, if you are wondering whether μEMA and EMA collect similar data. We compared the user-level variability captured by μEMA and EMA across 11 affect-based constructs. We found a moderate to strong +ve correlation between μEMA and EMA variability across constructs.

Third, when we measure user burden at the end of 12 months of data collection (only those who completed the study), μEMA was still perceived as less burdensome among those with possible survivor bias in data collection.

Second, we observed that regardless of the users' engagement with the data collection study (e.g., those who completed vs. withdrew vs. unenrolled), μEMA was consistently perceived as less burdensome, despite much higher interruption longitudinally.

This means, users who withdrew or were unenrolled because of poor engagement with EMA were twice as likely to answer μEMA surveys in real-world settings. We also suspect some survivor biases among those users who successfully completed 12 months of data collection.

First, we found that μEMA response rates were highest among those users who were unenrolled by research staff or voluntarily withdrew from data collection because of EMA burden. This response rate difference was negligible among those who completed 12 months of data collection.

As a result, we modeled ~1.3 million μEMA surveys and 14.9K EMA surveys collected across N = 177 participants, resulting in ~50K data collection days.

But for μEMA, each interruption presented only one micro-question with a yes/no type answer that can be responded to with a quick micro-interaction (taking hardly 2 seconds). In EMA, users answered long surveys with multiple back-to-back questions.

We used data collected in the TIME study, where users responded to surveys using μEMA and EMA in a burst-based longitudinal experiment. The μEMA method collected data on a smartwatch 4 times/hr for ~270 days. The EMA method collected data once/hr for ~90 days.

Most of my research involved personalizing the experience sampling methods. We proposed a novel method of using smartwatch interactions to collect self report at scale. One paper showed how smartwatches may be less prone to non response bosses in longitudinal studies. dl.acm.org/doi/abs/10.1...

Also given that data was collected in early adoption of the platform, will it change the way people use it?

I’m finding this starter pack directory useful to rebuild network. blueskydirectory.com

I’m treating the starter pack labels as a broad intent search query and so posts that are not relevant from the child accounts can be automatically hidden from my timeline. Just thinking aloud

Yeah that would be nice. Right now those tabs are manual. But also, let’s say I follow “HCI researchers” starter pack. I’m following folks with the expectation that it’ll be about relevant HCI research. In a perfect world, I’ll happy to not see non-HCI posts from those accounts

Also wondering hope these EMA surveys are delivered. On a mobile phone, numeric rating scales tend to be below the field biasing responses. Usually fully labeled 5 point scale has worked better for us. Of course you compromise on sensitivity

Hey Chris! Fellow Quant UXR here! Can you add me to the pack? 🙏