Baking a cake with a broken oven thermometer doesn't just give you a bad cake - it gives you confidently wrong information about what temperature works. You adjust your recipe based on faulty readings, blame the flour, maybe curse your altitude, and never suspect the thermometer itself. That's essentially what happens when researchers collect health data from wearable devices and don't bother asking why some of the data decided to go AWOL.

A new study published in the Journal of Medical Internet Research tackled exactly this problem, and its setting makes it even more fascinating: the informal settlements of Kampala, Uganda.

Smartwatches in the Slums

Here's the setup. Three hundred women living in slum communities in Kampala were given Garmin smartwatches as part of The Onward Project on Well-being and Adversity. The watches tracked sleep data over five nights. Simple enough, right? Strap on a watch, go to sleep, wake up, science happens.

Except roughly 30% of the nighttime data just... vanished.

Now, if you've ever worn a fitness tracker to bed, you know the drill. You charge it and forget to put it back on. You rip it off at 2 AM because it's uncomfortable. Your toddler uses it as a chew toy. The reasons data goes missing are as varied as the people wearing the devices. But here's where it gets interesting: the reason data disappears matters enormously for whether your study's conclusions are worth anything.

The Three Flavors of Missing Data

Statisticians love categorizing things, and missing data is no exception. There are three main "mechanisms" of missingness, and they have names that sound like increasingly bad dating scenarios:

Missing Completely at Random (MCAR): The data gaps have absolutely nothing to do with anything in the study. A cosmic coin flip. Your watch battery died because batteries die sometimes. This is the easiest type to handle statistically because you can basically just ignore the gaps and carry on.

Missing at Random (MAR): The missingness is related to something you did measure. Maybe younger participants were more likely to take the watch off. You can account for this because you have age data.

Missing Not at Random (MNAR): The missingness is related to the very thing you're trying to measure. If people with terrible sleep are the ones removing the watch because their sleep is terrible, you've got a real problem. Your dataset now systematically excludes the people you most need to hear from.

This is exactly why the Kampala researchers didn't just count the blank cells in their spreadsheet and shrug. They deployed four different analytical methods to figure out which flavor of missingness they were dealing with.

CSI: Missing Data Edition

The research team brought out the full forensic toolkit. First, classic pattern analysis - just looking at where the holes were and whether they clustered in non-random ways. Then the Little test, a statistical test specifically designed to sniff out whether data is MCAR. And finally, two machine learning approaches: a random forest model and a logistic regression model.

The results were unambiguous. The data was not missing completely at random. The Little test came back with a p-value less than 0.001, which in statistics is basically screaming "THIS IS NOT A COINCIDENCE" through a megaphone.

Three dominant patterns emerged among the participants' data: some women had complete data across all five nights (the gold stars), some had data missing only on the fifth and final night (maybe watch fatigue?), and some had data missing across all five nights (suggesting they may never have worn the device to sleep at all, or the device never successfully captured nighttime data).

Both machine learning models agreed with the traditional approaches, which is reassuring. When a random forest and a logistic regression are nodding along together, you can be reasonably confident in the finding.

Why This Matters Way Beyond Uganda

Wearable fitness trackers have been hailed as a potential game-changer for public health research, especially in low-resource settings where traditional clinical monitoring infrastructure is sparse or nonexistent. The appeal is obvious: they're relatively cheap, they collect data continuously, and they don't require a nurse hovering over someone with a clipboard.

But the promise of wearables hinges entirely on the quality of the data they collect. And this study highlights a tension that the global health community needs to wrestle with. The very environments where wearables could be most transformative - places with limited healthcare infrastructure - are also the environments where logistical barriers to consistent data collection are highest.

Think about it. Reliable electricity for charging isn't guaranteed. The cultural context of wearing an unfamiliar device to bed may create discomfort. Environmental conditions might affect device performance. And when participants live in challenging circumstances, asking them to maintain a consistent relationship with a piece of consumer electronics for five consecutive nights is asking more than it sounds.

The Bigger Lesson: Don't Just Plug the Holes

What's genuinely useful about this study is that it doesn't just wave its hands and say "missing data is bad." It provides a methodological roadmap. The combination of traditional statistical tests and machine learning classification models offers researchers a practical framework for diagnosing why their data has gaps before they decide how to handle them.

This matters because the wrong approach to missing data can introduce bias that's invisible in the final results. If you assume data is MCAR when it's actually MNAR, your statistical corrections will give you neat, confident numbers that are quietly wrong. You're back to the broken oven thermometer.

The study also makes a compelling case for using multiple methods to assess missingness mechanisms, rather than relying on a single test. The agreement between the pattern analysis, Little test, random forest, and logistic regression gives much stronger evidence than any single approach alone.

Looking Forward

As wearable technology continues to infiltrate health research globally - and it will, because the economics are too compelling to ignore - studies like this one lay the groundwork for doing it responsibly. The data from your Garmin or Fitbit isn't automatically trustworthy just because a computer collected it. The gaps in the data tell a story too, and ignoring that story means missing something that might be more informative than the data itself.

Three hundred women in Kampala wore smartwatches to bed for five nights. What they taught researchers about the absence of data may end up being as valuable as any sleep metric they recorded.

This blog post discusses research findings and should not be taken as medical advice. If you have concerns about sleep health or wearable device accuracy, please consult a healthcare provider. Research discussed here represents ongoing scientific investigation and clinical validation is still in progress.

All images used in this post are decorative illustrations only and do not represent or reflect the accuracy, reality, or correctness of the referenced research.

Primary Source: Determining a Likely Mechanism of Missingness in Repeated Measures Sleep Data From Wearable Fitness Trackers: Longitudinal Analysis. Journal of Medical Internet Research. 2025. PubMed: 41915671