Abstract
Based on technological advances in sensing modalities, randomized trials with primary outcomes represented as high-dimensional vectors have become increasingly prevalent. For example, these outcomes could be week-long time-series data from wearable devices or high-dimensional neuroimaging data, such as from functional magnetic resonance imaging. This paper focuses on randomized treatment studies with such high-dimensional outcomes characterized by sparse treatment effects, where interventions may influence a small number of dimensions, e.g., small temporal windows or specific brain regions. Conventional practices, such as using fixed, low-dimensional summaries of the outcomes, result in significantly reduced power for detecting treatment effects. To address this limitation, we propose a procedure that involves subset selection followed by inference. Specifically, given a potentially large set of outcome summaries, we identify the subset that captures treatment effects, which requires only one call to the Lasso, and subsequently conduct inference on the selected subset. Via theoretical analysis as well as simulations, we demonstrate that our method asymptotically selects the correct subset and increases statistical power.
Introduction: The Challenge of High-Dimensional Outcomes
Modern randomized controlled trials (RCTs) are increasingly high-dimensional. Instead of measuring a single outcome, researchers collect thousands of data points per participant, such as continuous glucose monitor (CGM) traces or neural activity maps.
The primary challenge is that a treatment might only affect a small portion of that data—a "sparse" effect. If a researcher pre-commits to a broad summary, such as an average over 24 hours, they risk washing out a significant but localized effect. Conversely, testing every single dimension individually leads to a severe statistical penalty due to multiple testing corrections, reducing the power to detect any effect at all.
Proposed Methodology
The authors propose a two-stage procedure to bridge this gap, allowing the data to identify where the effect lies without compromising statistical validity.
1. Subset Selection via the "Lasso Trick"
The core innovation is casting the problem of finding the subset of dimensions that maximizes the Hotelling $T^2$ statistic as a weighted linear regression problem. By applying Lasso ($\ell_1$ regularization) to this weighted regression, the method efficiently identifies the most relevant dimensions even when the number of potential summaries ($p$) exceeds the number of participants ($n$). This approach considers dimensions jointly, which is more effective in low signal-to-noise environments than ranking dimensions individually.
2. Inference via Sample Splitting
To ensure valid inference and avoid "p-hacking," the researchers employ sample splitting. The dimensions identified in the first stage are tested on a second, independent split of the data using standard causal inference estimators (e.g., difference-in-means or covariate-adjusted estimators). The method also allows for the inclusion of pre-treatment covariates to further reduce noise.
Experimental Validation
The method was validated using both synthetic simulations and semi-synthetic data derived from the 4T study at Stanford.
In the CGM experiments, 5,000 real-world glucose traces from 149 patients with Type 1 diabetes were used. Synthetic treatment effects (e.g., a 5-unit drop) were added at random two-hour intervals. The proposed method consistently achieved higher statistical power than standard approaches using fixed windows and multiple testing corrections, particularly when the exact timing of the intervention effect was unknown.
Key Results
- Increased Power: The Lasso-based selection reliably recovered sparse effects that were otherwise lost in broad averages.
- Superior Recovery: In simulations with 500 outcome dimensions, the method identified "active" dimensions more accurately than baseline ranking methods, especially in small sample sizes ($n=200$).
- Theoretical Consistency: The authors proved that as sample size grows, the method asymptotically selects the correct subset of dimensions.
Discussion and Future Work
While the method significantly boosts power, it currently relies on sample splitting, which reduces the data available for the final estimation. Future work may explore "multi-sample splitting" to mitigate this. Additionally, the method is optimized for sparse effects; in cases where a treatment has a small effect across all dimensions (a "dense" effect), alternative approaches may be more suitable. Finally, developing robust ways to aggregate subsets across multiple data splits remains an open research area.
Summary updated