arXiv:2610.41226v1 Announce Type: new Abstract: Long-term, diverse personal informatics data enables athletes to understand and act on their health and performance goals, whether assessing readiness right now, tracking how soreness relates to training load over time, or predicting how sleep affects next-day performance. However, the inherent missingness in longitudinal data, if handled inappropriately, can bias these analyses. To our knowledge, no prior work has systematically benchmarked classical and deep learning imputation methods on long-term self-tracking data from athlete populations. We evaluate six imputation methods across four missingness mechanisms at missingness rates spanning 10-80%. Using seven diverse long-term self-tracking datasets from 231 individuals, we assess both point accuracy and the preservation of temporal and multivariate structure. Our results show that no single method dominates across all conditions and reveal a critical trade-off: while PAIN is superior on point-wise accuracy, deep learning methods (BRITS, GRU-D, and GRIN) are better at preserving multivariate and temporal structure, particularly under higher missingness. Our findings offer a practical, goal-conditioned framework that guides researchers in selecting imputation strategies aligned with both the underlying missingness mechanism and the specific analytical goal of long-term self-tracking studies.