--- title: "Auditing Decision Drift Over Time" author: "Subir Hait, Michigan State University" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Auditing Decision Drift Over Time} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r setup, include = FALSE} knitr::opts_chunk$set(collapse = TRUE, comment = "#>") ``` > **Disclaimer:** `AIGovernance` is a statistical and documentation support > tool. It does not provide legal advice or certify compliance. ## Why a single-period audit is not enough An employer that selects 60% of applicants in two successive years may look stable. The stable rate can hide a rule that became more lenient for some applicants and stricter for others. A falling rate may come from a weaker applicant pool while the rule stays the same. Or the rate may change because cases are routed to different recruiters. The four-fifths rule and the NYC Local Law 144 impact ratio compare groups within one period, so they cannot separate these explanations. `aigov_drift()` compares two periods and reports three things: 1. **A decomposition** of the change in each decision rate into case mix, allocation of cases to decision units, and within-unit rule change. 2. **Net drift** $\Delta_g$: the average change in the decision rule for comparable cases, over a fixed reference case mix. 3. **Absolute drift** $A_g$: the average size of the change in the rule, so that increases and decreases do not cancel. A large $A_g$ with a small $\Delta_g$ is **masked drift**. ## Example data `hiring_panel` contains 4,500 simulated screening decisions by 24 recruiters over three years. It was built with known mechanisms: weaker applicant pools, offsetting rule changes for applicants with and without a degree, routing of strong applicants to lenient recruiters, and a higher bar for "advance". ```{r data} library(AIGovernance) data(hiring_panel) round(with(hiring_panel, tapply(selected, list(group, period), mean)), 3) ``` Selection fell for every group. The question is why. ## Single-period audits ```{r single} gov <- aigov_build(hiring_panel[hiring_panel$period == 2026, ], selected, group, ref_group = "A") gov <- aigov_adverse_impact(gov) ``` The impact ratios describe 2026 alone. They say nothing about what changed. ## Drift audit ```{r drift} gov <- aigov_build(hiring_panel, selected, group, ref_group = "A") gov <- aigov_drift(gov, period = period, covariates = c("qualification", "degree"), unit = recruiter, baseline = 2024, current = 2026, B = 50, seed = 1) ``` Read the output in this order: * **Decomposition.** Selection fell by about 18 points. About half of the fall is case mix (weaker applicants) and about half is allocation (weaker applicants were routed to stricter recruiters). The within-unit rule change is near zero: on average, recruiters did not become stricter. * **Drift.** Net drift for comparable applicants is about -8 points. This is a change in the *system's* treatment of comparable cases, and the decomposition shows that it comes from routing, not from recruiters. Absolute drift is about twice the net drift, so the rule also moved in opposite directions for different applicants (by construction, applicants with a degree gained and those without lost). * **Units.** The noise-corrected between-unit standard deviation is small and no recruiter is flagged, so recruiters changed by similar amounts. Because the data are simulated, the truth is known: ```{r truth} tr <- attr(hiring_panel, "truth") tr$decomposition tr$drift[, c("group", "true_net_drift", "true_abs_drift")] ``` The estimates are close to the truth. The net drift estimate is about two standard errors from its true value of -0.05, which is ordinary sampling error at 1,500 cases per period. Repeated simulation with `aigov_simulate_drift()` shows no systematic bias for this design. ## Categorical decisions The binary outcome `selected` only records passing the first screen. With the three-category decision the audit also sees movement between "review" and "advance". ```{r categorical} gov <- aigov_drift(gov, period = period, covariates = c("qualification", "degree"), decision = decision, focal = "advance", baseline = 2024, current = 2026, B = 50, seed = 1, verbose = FALSE) d <- gov$results$drift$decomposition d[d$group == "Overall", c("category", "raw_change", "case_mix", "rule_change", "net_drift")] ``` The "reject" rate rose and both "review" and "advance" fell. A binary audit of `selected` would miss the extra movement from "advance" to "review" for applicants who still pass the first screen. ## Monitoring several periods ```{r monitor} gov <- aigov_monitor(gov, period = period, covariates = c("qualification", "degree"), B = 50, seed = 1) ``` Each later period is compared with the baseline, and net-drift p-values are adjusted for the number of comparisons. In 2025 the net change is within tolerance but the absolute change is not: this is masked drift. By 2026 the absolute change is larger still. The net change is near the tolerance, and after adjustment for the number of comparisons it is not clearly beyond it, so those rows are reported as conditional drift. This is valid for prespecified comparisons. Repeated looks at accumulating data need sequential methods. ## Evaluating the audit by simulation `aigov_simulate_drift()` generates data under known mechanisms, so the false-alarm rate and power of any audit can be studied. For example, a pure change in case mix should not raise an alert: ```{r sim} sim <- aigov_simulate_drift(n = 1500, scenario = "case_mix", seed = 3) g <- aigov_build(sim, selected, group, ref_group = "A") g <- aigov_drift(g, period = period, covariates = c("qualification", "degree"), B = 50, seed = 1, verbose = FALSE) g$results$drift$drift[, c("group", "net_drift", "abs_drift_bc", "alert")] ``` ## What the audit does not establish Drift is a change in decisions conditional on the recorded characteristics. It does not identify the causal effect of an AI system or of human discretion. A change in an unrecorded, decision-relevant characteristic appears as drift (scenario `"unobserved"` in the simulator). An alert is a reason to review rule changes, data coding, model versions and override practice. It is not a finding of harm. ## References Kitagawa, E. M. (1955). Components of a difference between two rates. *Journal of the American Statistical Association*, 50(272), 1168--1194. Cai, T. T., Namkoong, H., & Yadlowsky, S. (2023). Diagnosing model performance under distribution shift. arXiv:2303.02011.