Determining a Likely Mechanism of Missingness in Repeated Measures Sleep Data From Wearable Fitness Trackers: Longitudinal Analysis.

Background: Wearable fitness trackers have become a valuable tool in public health research due to their ability to collect large-scale, individual-level data at a low cost. Because of this, wearables have the potential to mitigate the urgent need for health research in low-resource settings. However, their use has been largely limited to high-income settings, and a major challenge remains: high rates of missing data. This problem may be exacerbated in low-resource environments where logistical and operational barriers further complicate data collection. Wearable sleep data collected in an urban informal settlement in Uganda during The Onward Project on Well-being and Adversity project revealed substantial challenges related to missingness in low-resource research settings. Objective: The study aimed to characterize the patterns and frequency of missing data and to demonstrate the use of analytical methods to determine the underlying missingness mechanism. Methods: For this study, 300 women in slum communities in Kampala were equipped with Garmin smartwatches that collected data over 5 days. The nature of missingness was assessed through four methods: pattern analysis, the Little test, a random forest classification model, and a logistic regression classification model. Pattern analysis is an established method of identifying missing data patterns, while the Little test is used specifically to identify missing completely at random data. Random forest and logistic regression models are more recent methods proposed to identify the mechanism of missingness; both were used to examine agreement between the 2 methods and to ensure a thorough examination of missingness patterns. Results: Approximately 30% of nighttime data were missing. Three patterns were identified that occurred in over 10% of participants' data: no data missing, fifth night of data missing, and all nights of data missing. Pattern analysis and the Little test (P<.001) indicated that the data were not missing completely at random. Both the random forest (area under the curve=0.7) and logistic regression models suggested that the data were missing at random (MAR). Conclusions: Evidence of missingness in this dataset was consistent with MAR. Potential causes of missingness include device removal, battery failure, and technical malfunctions. These findings have important implications for both wearable device users and future research. Understanding the mechanisms behind missingness can inform strategies to improve data quality, particularly in low-resource settings, and allow researchers to manage missingness using appropriate strategies designed to handle MAR data without introducing bias into the results.