JMIR Publications

ABSTRACT

Background:

Mobile sensing promises scalable, ecologically valid mental health measurement by deriving behavioral features from smartphones. This promise rests on an untested assumption: that these features reflect the person carrying the device rather than the device itself. iOS and Android differ sharply in how they govern background sensing and logging, so the operating system may systematically distort the features, and because platform ownership correlates with income, age, and region, such distortion would act as a measurement confounder.

Objective:

This study aimed to determine whether the smartphone platform operates as a measurement confounder in mobile sensing for mental health, by characterizing the magnitude, localization, temporal stability, removability, and psychological independence of platform-induced differences in behavioral features.

Methods:

We analyzed the GLOBEM datasets, comprising four annual cohorts of college students (INS-W_1 to INS-W_4; 155, 218, 137, and 195 participants; 705 person-years) assessed with an identical passive-sensing pipeline. From a pre-specified panel of 42 all-day features across seven sensor domains, we (1) compared iOS and Android participants per feature using Cliff's δ with Benjamini-Hochberg correction and combined cohorts by random-effects meta-analysis; (2) trained L2-regularized logistic-regression classifiers to predict platform under participant-grouped cross-validation, then localized the signal by leave-out analysis and tested whether normalization or discretization removed it; and (3) regressed the most biased features on platform while adjusting for seven validated psychological instruments.

Results:

Platform differences were large, reproduced in direction across all four years, and were concentrated in connectivity and communication features (connectivity median absolute δ 0.36 versus behavioral 0.09; Mann-Whitney U=82, P=.015). Bluetooth scan counts were the most extreme (pooled δ −0.99, 95% CI −1.00 to −0.98), with median Android counts 18 to 112 times higher than iOS. The feature set predicted platform almost perfectly (area under the receiver-operating-characteristic curve [AUROC] 0.99 to 1.00 in the three larger cohorts), and leave-out analysis localized this to Bluetooth and call features, with behavioral-only AUROC falling to chance (0.46 to 0.67). Standard transformations were insufficient: discretization left predictability essentially intact (AUROC 0.987) and per-participant normalization reduced it only partially (0.993 to 0.946), acting specifically on connectivity features. The platform effect persisted after psychological adjustment (Bluetooth iOS β −2.03, 95% CI −2.13 to −1.93, P<.001, n=695; call iOS β −1.60, P<.001), with all psychological covariates near zero.

Conclusions:

Within one multi-year dataset, the smartphone platform operates as a partially irreducible measurement confounder, concentrated in connectivity and communication features, not removed by routine preprocessing, and not explained by measured psychology. The device should be modeled explicitly to support valid, reproducible, and equitable mobile-sensing research. Clinical Trial: Not Applicable