Selection-Corrected Statistical Inference for Region Detection With High-Throughput Assays.

Area of Science:

• Biostatistics
• Genomics
• Neuroimaging

Background:

• High-dimensional measurement assays detect signals in spatially organized data, like DNA methylation and brain activity.
• Common detection methods group neighboring sites above a threshold, but this introduces selection bias.
• Generalizing findings to a population requires accounting for this bias.

Purpose of the Study:

• Introduce pin-down inference, a novel model and framework for population inference in detected regions.
• Provide non-asymptotic estimators for mean effects that correct for local selection bias.
• Accommodate non-stationary covariances for robust region comparison.

Main Methods:

• Developed a pin-down inference framework for detected regions.
• Utilized conditional one-parameter exponential family with truncations matching selection constraints.
• Implemented a secondary screening-and-adjustment step to control false coverage rate.

Main Results:

• Pin-down inference provides accurate point and confidence interval estimators for mean effects.
• The method accounts for local selection bias and non-stationary covariances.
• Demonstrated superior power compared to existing conditional and non-parametric approaches in DNA methylation data analysis.

Conclusions:

• Pin-down inference offers a robust solution for population inference in spatially organized high-dimensional data.
• The framework effectively addresses selection bias and complex covariance structures.
• Enables more reliable comparisons of detected regions across various scientific domains.