Title:	Longitudinal Bias Auditing for Sequential Decision Systems
Version:	0.1.0
Description:	Provides tools for detecting, quantifying, and visualizing algorithmic bias as a longitudinal process in repeated decision systems. Existing fairness metrics treat bias as a single-period snapshot; this package operationalizes the view that bias in sequential systems must be measured over time. Implements group-specific decision-rate trajectories, standardized disparity measures analogous to the standardized mean difference (Cohen, 1988, ISBN:0-8058-0283-5), cumulative bias burden, Markov-based transition disparity (recovery and retention gaps), and a dynamic amplification index that quantifies whether prior decisions compound current group inequality. The amplification framework extends longitudinal causal inference ideas from Robins (1986) <doi:10.1016/0270-0255(86)90088-6> and the sequential decision-process perspective in the fairness literature (see https://fairmlbook.org) to the audit setting. Covariate-adjusted trajectories are estimated via logistic regression, generalized additive models (Wood, 2017, <doi:10.1201/9781315370279>), or generalized linear mixed models (Bates, 2015, <doi:10.18637/jss.v067.i01>). Uncertainty quantification uses the cluster bootstrap (Cameron, 2008, <doi:10.1162/rest.90.3.414>).
License:	MIT + file LICENSE
Encoding:	UTF-8
Language:	en-US
RoxygenNote:	7.3.3
Imports:	dplyr (≥ 1.1.0), tidyr (≥ 1.3.0), ggplot2 (≥ 3.4.0), rlang (≥ 1.1.0), cli (≥ 3.6.0), purrr (≥ 1.0.0), tibble (≥ 3.2.0)
Suggests:	mgcv, lme4, boot, knitr, rmarkdown, testthat (≥ 3.0.0)
Config/testthat/edition:	3
VignetteBuilder:	knitr
LazyData:	true
Depends:	R (≥ 4.1.0)
URL:	https://github.com/causalfragility-lab/AIBias
BugReports:	https://github.com/causalfragility-lab/AIBias/issues
NeedsCompilation:	no
Packaged:	2026-03-31 21:42:48 UTC; Subir
Author:	Subir Hait [aut, cre]
Maintainer:	Subir Hait <haitsubi@msu.edu>
Repository:	CRAN
Date/Publication:	2026-04-04 09:30:08 UTC

AIBias: Longitudinal Bias Auditing for Sequential Decision Systems

Description

AIBias treats algorithmic bias as a longitudinal process on repeated decisions, not a one-time snapshot. The package provides tools for:

• Trajectory analysis: How group-level decision disparities evolve over time

• Transition analysis: Where compounding occurs in decision sequences

• Amplification detection: Whether prior decisions worsen later inequality

Core Workflow

# Build the audit object
obj <- aib_build(data, id = "unit_id", time = "wave", group = "group",
                 decision = "approved")

# Run the full audit pipeline
result <- aib_audit(obj, ref_group = "majority")

# Or step by step:
desc   <- aib_describe(obj)
trans  <- aib_transition(obj)
amp    <- aib_amplify(obj)

plot(obj, type = "trajectory")
summary(obj)

Author(s)

Maintainer: Subir Hait haitsubi@msu.edu (ORCID)

See Also

Useful links:

• https://github.com/causalfragility-lab/AIBias

• Report bugs at https://github.com/causalfragility-lab/AIBias/issues

Covariate-adjusted bias trajectories

Description

Estimates covariate-adjusted bias trajectories by fitting a model for Pr(D_{it} = 1 \mid X_{it}, A_i, t) and computing marginal predicted disparities by group and time.

Usage

aib_adjust(
  object,
  formula,
  method = c("glm", "gam", "mixed"),
  ref_group = NULL,
  verbose = TRUE
)

Arguments

Value

The aibias object with ⁠$adjusted⁠ populated, containing:

• trajectory: adjusted bias trajectory

• marginal_rates: marginal predicted rates by group and time

• model: the fitted model object

• formula_used: the full formula passed to the model

Examples

data(lending_panel)
obj <- aib_build(lending_panel, "applicant_id", "year", "race", "approved")
obj <- aib_adjust(obj, formula = ~ income + credit_score, method = "glm",
                  ref_group = "White")

Compute dynamic bias amplification indices

Description

Estimates the amplification index A_{g,r}(t) = B_{g,r}(t|1) - B_{g,r}(t|0), which measures how conditioning on prior decision state changes the group disparity at time t. Non-zero amplification indicates that prior decisions are shaping current disparities—the hallmark of compounding bias.

Usage

aib_amplify(object, ref_group = NULL, verbose = TRUE)

Arguments

Details

A decision system exhibits bias amplification if:

1. |B_{g,r}(t)| > |B_{g,r}(s)| for some t > s (disparity grows), AND

2. A_{g,r}(t) \neq 0 (prior decisions drive current disparity), OR

3. P_g(t) \neq P_r(t) (transition matrices are unequal)

Value

The aibias object with ⁠$amplification⁠ populated. Contains:

• lagged_disparity: B_{g,r}(t|d) for d equal to 0 or 1

• index: Amplification index A_{g,r}(t)

• cumulative: A_{g,r}(T) summed over time

• matrix_norm: \sum_t \|P_g(t) - P_r(t)\|

Examples

data(lending_panel)
obj <- aib_build(lending_panel, "applicant_id", "year", "race", "approved")
obj <- aib_transition(obj, ref_group = "White")
obj <- aib_amplify(obj, ref_group = "White")
obj$amplification$index

Run the full AIBias audit pipeline

Description

A convenience wrapper that runs the complete audit pipeline: aib_describe() → aib_transition() → aib_amplify(). Optionally runs aib_bootstrap() for uncertainty quantification.

Usage

aib_audit(
  object,
  ref_group = NULL,
  bootstrap = FALSE,
  B = 500,
  seed = NULL,
  verbose = TRUE,
  ...
)

Arguments

Value

A fully-populated aibias object.

Examples

data(lending_panel)
result <- aib_audit(lending_panel,
                    id        = "applicant_id",
                    time      = "year",
                    group     = "race",
                    decision  = "approved",
                    ref_group = "White")

summary(result)
plot(result, type = "trajectory")

Bootstrap confidence intervals for bias trajectories

Description

Computes bootstrap confidence intervals for the bias trajectory and cumulative burden by resampling units (cluster bootstrap).

Usage

aib_bootstrap(
  object,
  B = 500,
  conf = 0.95,
  ref_group = NULL,
  seed = NULL,
  verbose = TRUE
)

Arguments

Details

Uses the cluster (unit-level) bootstrap to preserve the panel structure. Units are resampled with replacement; all their time observations are retained.

Value

The aibias object with ⁠$bootstrap⁠ populated.

Examples

data(lending_panel)
obj <- aib_build(lending_panel, "applicant_id", "year", "race", "approved")
obj <- aib_describe(obj, ref_group = "White")
obj <- aib_bootstrap(obj, B = 200, seed = 42)

Build an AIBias audit object

Description

Constructs the core aibias S3 object from a panel dataset. Validates the panel structure and prepares internal data for downstream analysis.

Usage

aib_build(data, id, time, group, decision, outcome = NULL, verbose = TRUE)

Arguments

Details

The function expects a balanced or unbalanced panel where:

• id indexes units observed over multiple periods

• time is an ordered index (will be coerced to integer rank)

• group is a categorical variable indicating protected group membership

• decision is a binary 0/1 variable (1 = favorable decision)

Value

An object of class "aibias".

Examples

data(lending_panel)
obj <- aib_build(lending_panel,
                 id       = "applicant_id",
                 time     = "year",
                 group    = "race",
                 decision = "approved")

Describe longitudinal disparity trajectories

Description

Computes group decision rate trajectories, raw and standardized bias trajectories, and cumulative bias burden.

Usage

aib_describe(object, ref_group = NULL, weights = NULL, verbose = TRUE)

Arguments

Value

The aibias object with ⁠$bias⁠ populated. The bias element is a list with components:

• trajectory: raw bias trajectory B_{g,r}(t)

• trajectory_smd: standardized bias trajectory B^*_{g,r}(t)

• cumulative: cumulative bias burden CB_{g,r}(T)

• slope: first differences \Delta B_{g,r}(t)

• curvature: second differences \Delta^2 B_{g,r}(t)

• ref_group: reference group used

Examples

data(lending_panel)
obj <- aib_build(lending_panel, "applicant_id", "year", "race", "approved")
obj <- aib_describe(obj, ref_group = "White")
obj$bias$cumulative

Copy and run the paper figures script

Description

Copies the paper figures script to your working directory and optionally runs it. The script produces four publication-ready figures illustrating bias trajectory, transition asymmetry, amplification index, and cumulative burden from a toy simulation (N=20, T=3).

Usage

aib_figures(run = TRUE, dest = file.path(tempdir(), "paper_figures.R"))

Arguments

Value

The path to the copied script, invisibly.

Examples

# Copy and run immediately
aib_figures()

# Just copy to inspect first
aib_figures(run = FALSE, dest = file.path(tempdir(), "paper_figures.R"))

Check panel balance

Description

Check panel balance

Usage

aib_panel_info(object)

Arguments

Value

A tibble summarizing observation counts per unit.

Compute bias persistence above a threshold

Description

Compute bias persistence above a threshold

Usage

aib_persistence(object, threshold = 0.05)

Arguments

Value

A tibble with group-level persistence counts.

Compute group-specific decision transition matrices and disparities

Description

Estimates group-specific Markov transition probabilities p_g^{ab}(t) = Pr(D_{it} = b \mid D_{i,t-1} = a, A_i = g) and derives transition disparities, advantage retention, and recovery gaps.

Usage

aib_transition(object, ref_group = NULL, verbose = TRUE)

Arguments

Value

The aibias object with ⁠$transitions⁠ populated. Contains:

• probs: Transition probabilities by group and time

• pooled: Pooled transition probabilities (time-averaged)

• disparity: Transition disparity \Delta^{ab}_{g,r}(t)

• recovery_gap: Disparity in 0->1 transitions (recovery)

• retention_gap: Disparity in 1->1 transitions (retention)

• matrices: Named list of 2x2 transition matrices per group

Examples

data(lending_panel)
obj <- aib_build(lending_panel, "applicant_id", "year", "race", "approved")
obj <- aib_transition(obj, ref_group = "White")
obj$transitions$pooled

Synthetic Lending Panel Dataset

Description

A synthetic panel dataset simulating loan application decisions over six years for applicants from three racial groups. Designed to illustrate longitudinal bias analysis with AIBias.

The data are generated so that Black and Hispanic applicants face lower approval rates, lower recovery probabilities after denial, and lower retention probabilities after approval — producing compounding disparities over time.

Usage

lending_panel

Format

A data frame with 3,600 rows and 6 columns:

applicant_id

Character. Unique applicant identifier.

year

Integer. Year of application (2015–2020).

race

Factor. Racial group: White, Black, Hispanic.

income

Numeric. Annual income (thousands USD).

credit_score

Numeric. Credit score (300–850).

approved

Integer. Loan approval decision (1 = approved, 0 = denied).

Details

Transition parameters used in data generation:

| Group | P(approve | prev approved) | P(approve | prev denied) | |———-|————————|———————| | White | 0.82 | 0.65 | | Black | 0.62 | 0.38 | | Hispanic | 0.68 | 0.44 |

Source

Synthetic data generated via data-raw/lending_panel.R.

Examples

data(lending_panel)
head(lending_panel)
table(lending_panel$race, lending_panel$year)

Plot an aibias object

Description

Visualizes audit results. Supports four plot types:

• "trajectory": Bias trajectory B_{g,r}(t) over time

• "heatmap": Group-time disparity surface

• "transition": Group-specific transition probabilities

• "amplification": Amplification index A_{g,r}(t) over time

Usage

## S3 method for class 'aibias'
plot(
  x,
  type = c("trajectory", "heatmap", "transition", "amplification"),
  show_ci = TRUE,
  color_palette = NULL,
  ...
)

Arguments

Value

A ggplot2 object.

Examples

data(lending_panel)
obj <- aib_audit(lending_panel,
                 id = "applicant_id", time = "year",
                 group = "race",     decision = "approved",
                 ref_group = "White", verbose = FALSE)
plot(obj, type = "trajectory")
plot(obj, type = "heatmap")

Print an aibias object

Description

Print an aibias object

Usage

## S3 method for class 'aibias'
print(x, ...)

Arguments

Value

Invisibly returns x, called for its side effect of printing a concise summary of the audit object to the console.

Summarize an aibias object

Description

Produces a comprehensive audit summary including trajectory statistics, transition gaps, amplification indices, and narrative interpretation.

Usage

## S3 method for class 'aibias'
summary(object, ...)

Arguments

Value

Invisibly returns object, called for its side effect of printing a comprehensive audit summary to the console.