Maggie Wang
I am a final year PhD student in Biomedical Data Science at Stanford advised by Mike Baiocchi.
I do research in causal inference, focusing on developing new methods to design better randomized experiments. (See inspection-guided randomization and experimental evaluation of AI-assisted decision-making).
Outside of methods development research, I have been the lead statistician on several applied studies of health interventions and outcomes. (See quantifying under-referral rates for patients with chronic kidney disease, assessing the relationship between depression and substance use in adolescents and adults, and evaluating the effectiveness of a mental health chatbot for individuals with substance use disorder).
Recently, I have also been interested in evaluating the impact of frontier AI on education, healthcare, and the economy. I am a current member of the Stanford HAI AI Policy Working Group and previously worked as an intern with the AI Governance Lab at the Center for Democracy & Technology.
Before Stanford, I received a B.S. in Biomedical Engineering and in Computer Science from Johns Hopkins University.
Proposes a practical upper bound on how selection bias degrades a model’s performance on an unobserved target population using only partially observed selection mechanisms and target data, validated on synthetic, All of Us, and MIMIC-IV data.
Introduces effect estimands and a minimax stepped double wedge design for evaluating algorithm-assisted decision-making when human behavior adapts under repeated algorithm exposure.
A randomized trial found a Woebot-based digital intervention for substance use produced improvements comparable to an active psychoeducational control, with no significant difference between conditions.
Introduces inspection-guided randomization, a flexible restricted-randomization framework that filters out undesirable treatment assignments against pre-registered design criteria to improve experimental design.
A national survey of adolescents through adults found tobacco-and-cannabis co-use was associated with higher odds of screening positive for depression than using either substance alone.
Introduces Post-hoc Concept Bottleneck Models, which convert any pretrained neural network into an interpretable, concept-based model without sacrificing accuracy and support efficient global edits via concept-level feedback.
A national claims analysis found nearly half of chronic kidney disease patients at high risk of kidney failure had not seen a nephrologist within a year of that risk becoming identifiable.
* equal contribution