AIBias

Longitudinal Bias Auditing for Sequential Decision Systems

---

Why AIBias?

Standard fairness tools ask: "Is there a disparity at this moment?"

AIBias asks: "How does disparity evolve over repeated decisions, and are earlier decisions making it worse?"

In sequential systems — loan approvals, parole reviews, hiring pipelines, content moderation — a denial at time $t$ changes the features available at $t+1$. Static fairness metrics miss this compounding entirely.

AIBias treats algorithmic bias as a longitudinal process. It implements three pillars of dynamic bias analysis:

Pillar	Question	Key Estimand
Trajectory	How does disparity evolve over time?	$B_{g,r}(t)$, $CB_{g,r}(T)$
Transition	Where does compounding occur?	$\Delta^{01}{g,r}$, $\Delta^{11}{g,r}$
Amplification	Do prior decisions worsen later inequality?	$A_{g,r}(t)$

---

Installation

# Install from GitHub
remotes::install_github("causalfragility-lab/AIBias")

Once accepted on CRAN:

install.packages("AIBias")

---

Quick Start

library(AIBias)
data(lending_panel)

# One-shot full audit
result <- aib_audit(
  lending_panel,
  id        = "applicant_id",
  time      = "year",
  group     = "race",
  decision  = "approved",
  ref_group = "White"
)

summary(result)
plot(result, type = "trajectory")
plot(result, type = "heatmap")
plot(result, type = "transition")
plot(result, type = "amplification")

---

Step-by-Step Workflow

library(AIBias)
data(lending_panel)

# 1. Build the audit object
obj <- aib_build(lending_panel,
                 id       = "applicant_id",
                 time     = "year",
                 group    = "race",
                 decision = "approved")

# 2. Bias trajectories and cumulative burden
obj <- aib_describe(obj, ref_group = "White")
obj$bias$cumulative

# 3. Transition matrices, recovery and retention gaps
obj <- aib_transition(obj, ref_group = "White")
obj$transitions$recovery_gap
obj$transitions$retention_gap

# 4. Amplification index
obj <- aib_amplify(obj, ref_group = "White")
obj$amplification$cumulative

# 5. Covariate-adjusted trajectories
obj <- aib_adjust(obj,
                  formula   = ~ income + credit_score,
                  method    = "glm",
                  ref_group = "White")

# 6. Bootstrap confidence intervals
obj <- aib_bootstrap(obj, B = 500, seed = 42)

# 7. Full summary
summary(obj)

---

Paper Figures (Toy Simulation)

Reproduce the four methods paper figures using the built-in demo:

library(AIBias)
demo("paper_figures", package = "AIBias")

# or equivalently:
aib_figures()

This runs a toy simulation (N = 20, T = 3, 2 groups) and displays:

• Figure 1 -- Bias trajectory with 95% bootstrap CI
• Figure 2 -- Group-time disparity heatmap
• Figure 3 -- Transition probabilities (recovery vs retention)
• Figure 4 -- Amplification index over time

---

Core Estimands

Bias Trajectory

$$B_{g,r}(t) = \pi_g(t) - \pi_r(t)$$

The raw group disparity at each time point.

Cumulative Bias Burden

$$CB_{g,r}(T) = \frac{1}{\sum_t w_t} \sum_{t=1}^{T} w_t B_{g,r}(t)$$

Average disparity experienced across the full audit horizon.

Recovery and Retention Gaps

$$\Delta^{01}_{g,r} = Pr(D_t=1 \mid D_{t-1}=0, A=g) - Pr(D_t=1 \mid D_{t-1}=0, A=r)$$ $$\Delta^{11}_{g,r} = Pr(D_t=1 \mid D_{t-1}=1, A=g) - Pr(D_t=1 \mid D_{t-1}=1, A=r)$$

Measure whether disadvantaged groups recover from denials and retain approvals at lower rates than the reference group.

Amplification Index

$$A_{g,r}(t) = B_{g,r}(t \mid 1) - B_{g,r}(t \mid 0)$$

Non-zero amplification means prior decision state is actively modifying current group disparities.

---

Formal Definition: Bias Amplification

A decision system exhibits bias amplification for group $g$ relative to reference $r$ if:

1. $|B_{g,r}(t)| > |B_{g,r}(s)|$ for some $t > s$ (disparity grows over time), and
2. $A_{g,r}(t) \neq 0$ (prior decisions modulate current disparity), or $P_g(t) \neq P_r(t)$ (group transition matrices differ)

Proposition: If $p_g^{11}(t) < p_r^{11}(t)$ and $p_g^{01}(t) < p_r^{01}(t)$ for all $t$, then under common initial conditions the approval probability for group $g$ weakly decreases relative to group $r$ over time, producing nonnegative cumulative burden against group $g$.

This distinguishes static persistent bias (constant gap) from dynamic compounding bias (self-reinforcing gap driven by the decision process itself).

---

Package Structure

R/
  aib_build.R        # Constructor and validator
  aib_describe.R     # Bias trajectories and cumulative burden
  aib_transition.R   # Markov transition matrices and gaps
  aib_amplify.R      # Amplification index
  aib_adjust.R       # Covariate adjustment (GLM / GAM / GLMM)
  aib_bootstrap.R    # Cluster bootstrap confidence intervals
  aib_audit.R        # One-shot pipeline wrapper
  aib_figures.R      # Helper to access paper figure script
  plot.aibias.R      # ggplot2 visualizations
  summary.aibias.R   # print / summary S3 methods
  utils.R            # Internal helpers
  data.R             # Dataset documentation

demo/
  paper_figures.R    # Reproduces all 4 paper figures

inst/scripts/
  paper_figures.R    # Accessible via system.file() after install

---

Running Tests

devtools::test()
# Expected: FAIL 0 | WARN 0 | SKIP 0 | PASS 41

---

Citation

If you use AIBias in your research, please cite:

@software{Hait2026AIBias,
  title  = {AIBias: Longitudinal Bias Auditing for Sequential Decision Systems},
  author = {Hait, Subir},
  year   = {2026},
  url    = {https://github.com/causalfragility-lab/AIBias}
}

---

Author

Subir Hait
Michigan State University
haitsubi@msu.edu

---

License

MIT License. See LICENSE for details.