References

Pipeline Paper

From Clinical Trials to Real-World Impact: Introducing a Computational Framework to Detect Endpoint Bias in Opioid Use Disorder Research

Odom GJ, Brandt L, Marker A, Giorgi S, Jainarain G, Schwartz HA, Au L, Castro C, and the ENDPOINT Consortium. Drug and Alcohol Review. 2026;45(1):e70085.

View Paper

This is the primary publication describing the DAB Pipeline. It introduces the computational framework for detecting endpoint bias in OUD clinical trial data, demonstrates the pipeline across the CTN-0094 dataset, and presents the measurement invariance approach used to evaluate algorithmic fairness across demographic groups.

Foundational Paper

The DAB Pipeline is a direct extension of the following publication. The pipeline replicates and generalizes the modeling approach from this paper, applying it across a broader set of outcomes and demographic configurations to study algorithmic bias.

Individual-Level Risk Prediction of Return to Use During Opioid Use Disorder Treatment

Luo SX, Feaster DJ, Liu Y, et al. JAMA Psychiatry. 2024;81(1):45–56.

View Paper

What the paper did

Using data from four CTN trials (the same CTN-0094 dataset this pipeline is built on), Luo et al. developed and validated a multicenter decision-analytic model to predict individual-level return to opioid use during treatment. The model was trained on patient demographics and early treatment response features, and evaluated on a held-out test set.

What this pipeline extends

The DAB Pipeline takes the same modeling approach and asks a further question: does the model perform equally well across different demographic groups?

Specifically, the pipeline:

Replicates the core prediction task (return to use, abstinence, relapse events) across 8 outcome definitions derived from the same dataset
Introduces Propensity Score Matching to construct training cohorts with systematically varied majority/minority demographic compositions
Evaluates whether model performance is stable across those compositions — i.e., whether the outcomes are measurement invariant

If a model's performance degrades as the minority proportion of the training cohort increases, that is evidence of algorithmic bias in the original modeling approach.

Odom GJ et al. Measurement invariance framework for evaluating algorithmic bias in clinical prediction models. In preparation. 2025.

Provides the theoretical basis for the measurement invariance interpretation used in this pipeline.

Citing This Pipeline

If you use the DAB Pipeline in your research, please cite both the foundational paper and this pipeline:

@article{odom2026dar,
  author  = {Odom, Gabriel J. and Brandt, Laura and Marker, Aaron and Giorgi, Salvatore
             and Jainarain, Ganesh and Schwartz, H. Andrew and Au, Larry and Castro, Clinton
             and {the ENDPOINT Consortium}},
  title   = {From Clinical Trials to Real-World Impact: Introducing a Computational
             Framework to Detect Endpoint Bias in Opioid Use Disorder Research},
  journal = {Drug and Alcohol Review},
  year    = {2026},
  volume  = {45},
  number  = {1},
  pages   = {e70085},
  doi     = {10.1111/dar.70085}
}

@article{luo2024jama,
  author  = {Luo, Sean X. and Feaster, Daniel J. and Liu, Yan and others},
  title   = {Individual-Level Risk Prediction of Return to Use During Opioid Use Disorder Treatment},
  journal = {JAMA Psychiatry},
  year    = {2024},
  volume  = {81},
  number  = {1},
  pages   = {45--56},
  doi     = {10.1001/jamapsychiatry.2023.3596}
}

@software{jainarain2026dab,
  author    = {Jainarain, Ganesh and Marker, Aaron and Odom, Gabriel J. and Brandt, Laura},
  title     = {DAB Pipeline: Detecting Algorithmic Bias in OUD Treatment Data},
  year      = {2026},
  publisher = {GitHub},
  url       = {https://github.com/CTN-0094/Pipeline}
}