Research Roadmap
Six open questions that build directly on our completed paper suite. Each addresses a structural gap in N-of-1 experimental methodology — compliance, confounding, washout, graph transfer, changepoint detection, and active causal discovery.
Active Causal Discovery for N-of-1 Trials
Core idea
P8 shows trial count is the binding constraint on causal graph recovery and uses random intervention pulses. An active design that selects which cause to intervene on next — to resolve the most uncertain edge — would maximize discovery quality under a fixed trial budget. The acquisition function targets information gain about the edge indicator rather than Fisher information about a scalar effect.
Connects to
Bayesian optimal experiment design (Chaloner & Verdinelli, 1995) applied to graph recovery rather than parameter estimation.
Compliance-Aware N-of-1 Design
Core idea
N-of-1 designs assume perfect compliance. In practice, compliance is intermittent — creating an IV setting where assignment is random but actual exposure varies. This paper addresses ITT vs LATE estimation under partial compliance and dynamic allocation adapted for non-compliance patterns. The design question: does measuring compliance change the optimal allocation rule?
Connects to
Instrumental variables, LATE estimation, and partial identification bounds applied to within-person experiments.
Sequential Changepoint Detection for Experiment Redesign
Core idea
When should an experiment be stopped, adapted, or restarted? Current stopping rules detect statistical significance. A causal changepoint framework detects structural changes in the DGP — tolerance build-up, life change, seasonal shift — and triggers experiment redesign rather than just termination. Connects to the MRT/JITAI literature and sequential change-point detection (Page, 1954; Tartakovsky et al., 2014).
Connects to
CUSUM / Page-Hinkley tests + causal validity criteria for structural breaks in behavioral time series.
Federated Causal Graph Transfer
Core idea
P8 pools evidence about edge scores. A stronger transfer form pools entire causal graphs: users in the same response cluster share their learned graph as a prior for new users, not just arm means. A graph-transfer kernel measures how much two users' causal structures should be correlated given observable similarity — accelerating cold-start causal learning for new users.
Connects to
Causal transportability (Pearl & Bareinboim, 2011) applied to personal behavioral causal models; federated learning.
Confounded N-of-1 Causal Discovery
Core idea
P8 assumes intervention traces are clean. But users choose when to experiment based on their state — they run focus experiments on already-productive days. This confounds the discovered causal graph: focus appears more causally effective than it is. Applying P7's sensitivity framework to the causal discovery problem produces sensitivity bounds on edge scores under worst-case hidden-confounder bias.
Connects to
Partial identification + sensitivity analysis applied to PC/FCI algorithms in a within-person observational setting.
Adaptive Washout Period Estimation
Core idea
The required washout period is DGP-specific — depending on carryover decay rate — but is currently fixed by protocol. Learning washout from early experiment data means estimating carryover decay and adapting washout dynamically. P10 identifies washout as a validity-gate-critical parameter; this paper closes the loop by making it data-adaptive using survival analysis on the post-treatment decay trajectory.
Connects to
Survival analysis / changepoint methods on post-treatment carryover; MRT optimal washout selection.
Related research pages