Smartphone Missingness as a Depression Biomarker: A Baseline-Controlled Re-analysis of StudentLife

BackgroundWhether gaps in smartphone passive-sensing data carry psychological signal -- beyond what a baseline self-report already provides -- is a recurring question in digital phenotyping that has rarely been tested with multiplicity control and cross-validation on the same cohort. ObjectiveTo test whether participant-level missingness in passive sensing and ecological momentary assessment (EMA) is an incremental predictor of depression beyond a single baseline self-report, in the canonical StudentLife cohort. MethodsExploratory re-analysis of the publicly available StudentLife dataset. The original 2014 publication reports n = 48 enrolled undergraduates; the public archive contains 59 distinct sensor-instrumented UIDs, but the additional 11 are sparse-data or PHQ-9-incomplete records and contribute nothing to the between-person analyses, which use only the 38 participants with paired pre-term and post-term Patient Health Questionnaire-9 (PHQ-9). We computed 89 participant-level missingness features from nine continuous sensor streams, five phone-activity logs and 27 EMA prompts, and evaluated them under leave-one-out cross-validation with nested-CV-tuned hyperparameters, cluster-bootstrap confidence intervals, an omnibus joint F-test, and Benjamini-Hochberg multiplicity control. ResultsA pre-term PHQ-9 baseline alone explained 59% of out-of-sample variance in post-term scores (n = 32; 95% cluster-bootstrap CI [0.22, 0.81]). Tuned regularized linear models trained on missingness alone reached only the cohort-mean baseline; adding missingness to pre PHQ-9 did not improve performance. The omnibus joint F-test of all nine continuous-stream missingness rates against post-term PHQ-9, adjusted for pre-term PHQ-9, was non-significant (F (9, 27) = 0.43, P = 0.91). No individual feature survived multiplicity correction. A separate within-person day-level analysis (2,186 person-days) yielded a small valence-specific prospective effect (r = +0.082, 95% CI [+0.011, +0.162]) opposite in direction to the withdrawal hypothesis. ConclusionIn this cohort, smartphone-data missingness did not add incremental predictive value beyond a single baseline PHQ-9. The analysis is exploratory and StudentLife-specific; it should not be read as evidence that missingness is never informative. Plain-language summaryMany studies use the gaps in someones smartphone data -- missing GPS readings, missed survey prompts, fewer phone interactions -- as a possible warning sign of depression. This re-analysis tested that idea on a widely used public dataset from a class of 48 college students. After accounting for each students depression score at the start of the term, the gaps in their phone data added no useful information about their depression score at the end of the term. The result is specific to this dataset and does not mean that smartphone gaps are never informative, but it shows that such claims need careful baseline comparisons. Key PointsO_ST_ABSQuestionC_ST_ABSIs smartphone-data missingness an incremental depression biomarker beyond a single baseline self-report? FindingsIn an exploratory re-analysis of the StudentLife cohort (38 participants with paired PHQ-9 scores), missingness features did not improve prediction of post-term PHQ-9 beyond pre-term PHQ-9 under leave-one-out cross-validation, nested-CV-tuned models, an omnibus joint F-test, or multiplicity-controlled univariate screens. MeaningSmartphone-data missingness should not be interpreted as a psychological signal in absence of baseline-controlled, cross-validated, multiplicity-aware evidence. The result is specific to a high-functioning undergraduate cohort over a ten-week term and requires replication.