Quantitative psychology · Machine learning · Digital mental health
Selected Work
My work connects three questions: are we measuring meaningful change, can we predict risk, and which interventions help? I develop the statistical methods, models, and experiments needed to answer them with behavioral data.
Measurement and prediction
Clinical prediction starts with the quality of the signal. My work combines passive sensing and self-report with psychometrics, adaptive assessment, and longitudinal models to study mental health as it changes in daily life.
EVIDENT: measuring rapid clinical change
I lead a multimodal behavioral-health program combining Empatica and Garmin wearables, Android and iOS passive sensing, video, and ecological momentary assessment (EMA). I designed the protocol and measurement stack, selected vendors and data platforms, and led IRB, legal, and data-security review.
Study design: a 1,500-person randomized trial and a 300-person observational cohort. These are planned study sizes; the program’s work spans measurement development and study operations, not yet a completed outcome evaluation.
Award and research roles in my CV (PDF)
From smartphone screenshots to risk signals
I processed smartphone screenshots with OCR and vision-language classification to detect suicide-related language and predict momentary suicidal ideation. I also built a fine-tuning and inference stack on a 15-GPU L40S cluster, using Qwen, LoRA/PEFT, Unsloth, vLLM, and SGLang.
The model-selection harness uses user-grouped cross-validation, a lockbox test set, and promotion gates. The screenshot work produced a prospective machine learning study in JMIR Mental Health; related studies examined nighttime phone use and screenshot-derived screen time as markers of suicide risk.
Vision-language model study · Related digital-phenotyping publications
Adaptive assessment and longitudinal data
My collaborative work includes a multidimensional computerized adaptive test for momentary suicide-risk assessment, and studies of survey burden, compliance, missingness, and within-person dynamics. These address a practical measurement question: how much useful information can we collect without imposing unnecessary burden?
Adaptive assessment: development and usability study
Experiments and adaptive interventions
My role. I led causal and personalization analyses of two factorial-trial cohorts (2⁴ and 2⁶ designs) and a nested micro-randomized trial (MRT). I estimated causal excursion effects, examined differences in treatment response using causal forests and policy trees, and tested whether findings transported across cohorts.
Decision delivered. The completed analyses produced deployment rules for message timing and targeting. For Phase 2, I designed a bandit-driven just-in-time adaptive intervention (JITAI) and its simulation study.
Related ongoing collaborations. My digital-intervention work also includes co-investigator roles on an NIH-funded study of human and digital support in a meditation app for depression and anxiety, and a Templeton-funded program on personalized well-being interventions. I am an MPI on a project developing just-in-time support for outpatient suicide care. These are separate projects with distinct research roles and stages.
Funding and collaboration details in my CV (PDF)
Reliable methods and research software
I develop and co-develop statistical software for problems that recur across behavioral research: selecting variables, modeling change over time, and identifying differences between people.
regsem
Regularization and cross-validation for structural equation models.
Source code →MplusTrees
Co-developed software for recursive partitioning with structural equation models fit in Mplus.
longRPart2
Recursive partitioning of linear and nonlinear mixed-effects models for longitudinal data.
Source code →
What makes a model result credible?
My methodological work examines inflated prediction performance in suicide machine learning and the effect of measurement error on ML conclusions. This work informs the checks I build into model evaluation, alongside the development of regularized latent-variable models.
Inflated prediction performance · Measurement and machine learning
More on longitudinal modeling and methodological foundations
Ecological momentary assessment and intensive longitudinal data. Much of my methodological work concerns the analysis of intensive longitudinal data collected from clinical populations — handling momentary missingness, zero inflation, continuous-time dynamics, and computerized adaptive testing for in-the-moment risk assessment.
Regularized structural equation modeling. Beginning with the original Regularized Structural Equation Modeling paper (2016) and the regsem R package, I have contributed to a line of work extending regularization (lasso, elastic net, stability selection) to latent-variable models, and to understanding how measurement quality shapes the conclusions of machine learning applied to psychological data.
Machine learning methodology in psychology. A second methodological strand examines the reliability of machine learning claims in clinical psychology — including a frequently-cited commentary in Clinical Psychological Science on inflated prediction performance in suicide risk modeling, and a paper in Perspectives on Psychological Science on the often-overlooked role of measurement error in machine learning applications.
I co-authored Machine Learning for Social and Behavioral Research (Guilford, 2023), recipient of the 2025 Barbara Byrne Award for Outstanding Book from the Society of Multivariate Experimental Psychology. I have also taught applied deep learning with Python and advanced machine learning through Statistical Horizons.
Get in touch
I am interested in industry data science and research roles in digital health, measurement, and experimentation. My technical work spans Python (PyTorch, scikit-learn, pandas), R, SQL, Git, and Linux/HPC.