Outlier detection for standardized tests

Screening a national exam for anomalous results is only actionable if each flag can be explained to the people who must act on it.

Schematic. A screening rule separates ordinary variation from the few results it ranks as anomalous. No data is shown.
Status
Conference presentation (unpublished)
Period
2020 — 2022
Themes
Fair & reliable predictive modeling

Research question

How can statistical and machine-learning methods support interpretable screening for anomalous standardized-test results?

Why it matters

A university selection exam allocates places, and a flagged result can trigger review of an individual test taker. Screening therefore carries consequences in both directions: missed anomalies undermine the exam, and unexplained flags are not defensible to the agency that has to act on them.

The challenge

Anomalies in test data are rare, heterogeneous and not labeled, which rules out supervised detection and rewards methods whose output a reviewer can read. A single score is needed, but it has to be traceable back to the evidence that produced it.

Approach

The project developed an outlier-detection protocol combining statistical discrepancy measures, clustering-based screening and outlier-detection models into one interpretable anomaly score per test taker. The methodology was validated and presented to DEMRE, the government agency responsible for administering the exam.

Status

Presented at the XIV Chilean Conference on Operations Research (OPTIMA 2021) at Universidad Católica del Maule in March 2022. This is an unpublished conference presentation, not a publication.

A ranked table of twenty test takers. Columns give scores for language, mathematics, history and science, and the final column gives the combined anomaly score used to rank them. Several rows are highlighted in colour.
Example screening output: test takers ranked by the combined anomaly score, shown alongside the per-subject scores behind it. Reproduced from the presentation — see the slides for the method

Formal outputs

  1. Conference presentation (unpublished) · 2022

    On the outlier detection for standardized tests

    N. Acevedo Villena, C. Thraves and M. Varas · XIV Chilean Conference on Operations Research (OPTIMA 2021) · Universidad Católica del Maule · Talca, Chile

    See it on Publications

All research