Advancing Mobile Health Methods: Geographic Imputation of Missing Activity Space Data from Ecological Momentary Assessment (EMA) GPS Positions

Introduction: The integration of Ecological Momentary Assessment (EMA) with geospatial technologies such as Global Positioning Systems (GPS) and Geographic Information Systems (GIS), termed geographic EMA (GEMA), has revolutionized researchers’ ability to collect data on individuals’ activity spaces – the routine places people go for work, leisure, and so on – with corresponding information on moods, behaviors, and social interactions, in real time and in individuals’ natural environments. This approach has facilitated new types of research on how individuals’ mobility, and the socioeconomic and built environments they are exposed to throughout their daily lives, affect substance use and other health behaviors and outcomes. One challenge, however, is that GEMA often results in substantial missing activity space location data, due to GPS signal interference from the atmosphere, buildings, and other features. As part of the Special Conference Theme related to Mobile Health, this research presents the development and comparison of geographic imputation methods for estimating the location of missing activity space data collected using GEMA.

Methods: We develop two geographic imputation techniques which we adapt for activity space data from hot deck and centroid imputation approaches initially developed for non-spatial data and geocoding imputation. To test these techniques, we use GEMA data from a previously published analysis of the effect of neighborhood disadvantage, captured at the U.S. Census Bureau tract level, on momentary psychological stress among a sample of 137 urban adolescents (Mennis et al., 2016, Drug and Alcohol Dependence, 165, pp. 288-292). We test models of both direct and moderated effects of disadvantage on stress. We investigate the impact of geographic imputation, as well as listwise deletion, on model results.

Results: We found that listwise deletion altered the magnitude, significance, and standard errors of the disadvantage coefficient, particularly in situations where half the observations contained missing location data, and when estimating the significance and standard errors of the moderated effect. These impacts were ameliorated to some extent by using geographic imputation, particularly in mitigating the inflation of the standard errors, though results were more variable regarding the tests of moderation. We found that the activity space centroid imputation technique consistently outperformed hot deck imputation, though the difference was relatively modest.

Conclusions: These geographic imputation techniques may be extended in future research by incorporating regression-based and multiple imputation approaches from the non-spatial imputation literature, as well as from conventional geographic imputation and spatial interpolation research which focus on non-activity space spatial data.