Abstract
Machine learning models in high-stakes domains like healthcare and criminal justice are often marketed as "predictive" tools. However, there is a fundamental ambiguity: is the model identifying an individual's future agency or merely reproducing historical patterns? This paper introduces "backward baselines," a statistical framework to disentangle these pathways. By comparing model performance against baselines derived from background context, researchers can audit whether a system provides new insights or simply automates systemic stereotyping.
Introduction
In high-stakes decision-making, predictive models are used to anticipate individual actions. Yet, high accuracy alone does not reveal if a model is "predicting the future" (based on individual effort) or "predicting the past" (leveraging background variables determined before the individual acted). If a model relies primarily on "backward" pathways—such as demographics or historical associations—its use for individual-level interventions becomes ethically and practically questionable. This research addresses the lack of formal tools to distinguish between these two predictive mechanisms.
The Backward Baseline Framework
The authors propose a framework centered on three key variables:
- $W$ (Context): Background variables determined before the individual's actions (e.g., place of birth, parental education).
- $X$ (Covariates): Individual-level data measured in the present (e.g., personal financial history).
- $Y$ (Outcome): The future event being predicted.
The relationship between these variables is modeled as a directed acyclic graph where context $W$ influences both the current covariates $X$ and the outcome $Y$, while $X$ also influences $Y$.
Proposed Baselines
- Backward Prediction Baseline ($g^*$): The statistically optimal way to predict $Y$ using only $W$. If a complex model $h(X)$ does not significantly outperform this baseline, it is essentially stereotyping based on context.
- Backward Rounding Baseline ($g^h$): Predicts the model’s output $h(X)$ using $W$. A key technical contribution is showing that this baseline can audit "black-box" models even when ground-truth labels are unavailable, as it recovers the optimal backward predictor for calibrated models.
Empirical Evaluation
The authors evaluate their framework using datasets where the temporal separation between features and outcomes is clear, avoiding the "temporal leakage" common in many ML benchmarks:
- Medical Expenditure Panel Survey (MEPS): Predicting annual healthcare utilization.
- Survey of Income and Program Participation (SIPP): A longitudinal census survey used to predict poverty levels.
- ProPublica COMPAS Dataset: Recidivism risk scores.
Key Results
- Healthcare & Poverty: Models trained only on demographic background variables (age, race, etc.) nearly matched the performance of full-feature models. This suggests that social determinants are the primary drivers of accuracy in these domains.
- Recidivism: In the COMPAS dataset, a single "past" feature—prior convictions—accounted for almost all predictive power. The model provided very little additional "forward" insight into individual behavior beyond this historical marker.
Discussion and Insights
The paper highlights that accuracy is insufficient for evaluating social impact. Two models with identical accuracy can have vastly different ethical implications: one might reward individual agency, while the other reinforces systemic bias.
A "pure backward predictor" renders individual targeting futile; one could achieve the same results by intervening on anyone within a demographic group. Furthermore, the ability to audit models without labels using the "Rounding Baseline" is a significant step for transparency, allowing third-party probes into proprietary systems. The paper also formalizes a link between "forward prediction" and "Demographic Parity," suggesting that fair models should focus on what an individual does rather than where they came from.
Conclusion
While the choice of what constitutes "background context" ($W$) remains a normative decision, backward baselines provide a rigorous method for auditing predictive systems. The authors suggest these baselines should become standard in the "report card" for any supervised learning model deployed in social settings to ensure we are not simply automating the past.
Summary updated