Skip to content

ipd

Inference on Predicted Data — valid statistical inference when an outcome has been imputed by an ML/AI prediction algorithm — an R package that implements several recent methods (PostPI, PSPA, PPI / PPI++, PSPS, Chen & Chen, etc.) for downstream regression with ML-predicted outcomes, behind a single unified ipd() wrapper.

Talk at GBCC2025

  • Speakers: Stephen Salerno, Jiacheng Miao, et al.
  • Day / session: Day 4 (Thu Jun 26, 2025) — Oral Session 5, Grace Auditorium (chair: Scott Cain)
  • Talk title: "Inference with predicted data – what do we do after we have machine learned everything?"
  • Companion workshop: Day 2 (Tue Jun 24) — "Inference Challenges and Corrections for AI-Predicted Outcomes"
  • Slides: TBD
  • Video: GBCC2025 playlist
  • Abstract / schedule: GBCC2025 program

What it does

When an outcome variable in a downstream regression is partially imputed by an ML model (e.g. predicted survival from a deep model, predicted phenotype from an LLM, predicted cell-type label from a foundation model), naive statistical inference on that mixed real-and-predicted outcome is biased and has wrong standard errors. ipd implements the recent statistical-correction methods — PostPI, Prediction-Powered Inference (PPI / PPI++), PSPA, PSPS, Chen & Chen — behind a single user-facing ipd() wrapper that takes the labeled data, the unlabeled (predicted) data, and a formula, and returns a properly-corrected fit with valid coverage. Custom print, summary, tidy, glance, and augment methods make the corrected fit drop in to a tidyverse-style modeling pipeline.

How it works

Core idea. ipd uses the small labeled subset (where both the true outcome and the ML prediction exist) to estimate the bias and residual variance of the predictor, then propagates that correction through the regression fitted on the much-larger unlabeled subset — so the prediction error is treated as a measurement-error problem rather than ignored.

Inputs / outputs. Input is a single dataset with a label column flagging which rows are "labeled" (true outcome observed) vs "unlabeled" (only the predicted outcome), a formula relating the outcome and predictions to covariates, and a method choice. Output is an ipd fit object with corrected point estimates, valid standard errors / confidence intervals, and tidy/glance/augment accessors.

Key innovation. A single wrapper that dispatches across the major post-prediction-inference families (PostPI bootstrap + analytic, PPI, PPI++, PPI-All, PSPA, Chen & Chen) so users compare methods without learning six different APIs — and the inference results are valid under the underlying paper's assumptions even when the prediction model itself is biased.

Parameters / API surface worth knowing. - formula — for Chen & Chen the form is Y - f ~ X1 (Y = true outcome, f = predictions, X = covariates); other methods use standard regression formulas. - method — one of "chen", "postpi_boot", "postpi_analytic", "ppi", "ppi_plusplus", "pspa". - model — "mean", "quantile", "ols", "logistic", or "poisson". - label — column name identifying labeled vs unlabeled rows.

Canonical example. From the getting-started vignette: ipd(Y - f ~ X1, method = "chen", model = "ols", data = df, label = "set_label"). The labeled rows (where Y is observed) calibrate the bias of predictor f; the much larger unlabeled portion (where only f exists) supplies the precision; the returned fit's CIs cover at the nominal rate even though most rows used a machine-predicted outcome.

Where it fits in the corpus

  • AACR 2026: axis = agentic AI / ML; valid inference on ML-predicted outcomes is one of the foundational methodological questions for any "AI-first" cancer-biology workflow
  • EuroBioC 2025: no direct counterpart on post-prediction inference
  • ISMB 2026: scaffold; expect overlap on PPI-family methods
  • Nextflow Summit 2026: thematic adjacency with the AI-for-pipelines storyline (validating model outputs in production pipelines)

Notes

This is one of the more directly-useful methods talks at GBCC2025 for any group doing downstream stats on ML-labeled data. The fact that it ships as both a workshop and an oral talk indicates that the organizers expected significant user demand — and the field as a whole (PPI from Angelopoulos et al., PostPI from Wang et al., PSPA from Miao et al.) has converged on this being a real problem.