skip to content

Predictive Coding Strategies to Identify Patterns of Discrimination

by Joshua Gilliland

A view of Everlaw software on a laptop computer, displaying an overview of assignments

If lawsuits are about people who have experienced a wrong, requests for production are one of the most important tools in Civil Procedure to get at the “truth” of what happened in a case. In Jones v. Standard Consulting & Standard Testing & Eng’g Co., a gender and age discrimination case, the Plaintiff was terminated by the Defendant and requested discovery on all employees who were supervised by the Defendant. The Plaintiff’s theory was that the Defendant treated female employees differently with regard to work hours. The Court agreed the Plaintiff had alleged a pattern of discrimination by the Defendant, and ordered a supplemental production that required the Defendant to review the files of 50 employees (Jones v. Standard Consulting & Standard Testing & Eng’g Co., 2018 U.S. Dist. LEXIS 23835, at *6-7 (W.D. Okla. Feb. 14, 2018).

Blog Image - Scales justice

Using Predictive Coding to Find Responsive Discovery

There are a few tested strategies a producing party can use to identify the responsive discovery in their data. A producing party in a similar situation to the above defendant could leverage both search terms and predictive coding to find responsive discovery. The likely responsive information would be in termination documentation of employees, work schedules, and possible communications with female employees to show discrimination or retaliation.

  1. Prior to starting document review, issue coding and a prediction model should be created for responsive discovery.

  2. With the codes and the predictive coding model in place, the model can be trained by reviewing a small subset of documents.

  3. Other reviewing attorneys can focus on termination notices given to employees through searches or even search term reports. Responsive records can be coded accordingly.

The above strategies can help identify responsive information for the review team to train the predictive coding system. Once a prediction model has had a sufficient amount of data reviewed, it can predict documents for attorney review.

Document Review with Predictive Coding

Courts recognize that responding parties are best suited to determine the most effective methodologies for producing discovery.* The issue, if there is a challenge to a production, is whether it is inadequate or manifestly unreasonable.

For attorneys conducting review with predictive coding, the goal is to have a reasonable process to comply with Federal Rule of Civil Procedure Rule 26(g). This requires validating the data that is being reviewed. For example, attorneys should conduct review of data that is predicted to be irrelevant, to confirm it is indeed irrelevant. Everlaw allows admins to check the historical performance of the prediction model, by comparing it to a set of “hold out” data that is used for validation. This allows attorneys to see how scores change over time, providing insight on the performance of the prediction model and review.

Document review attorneys should provide feedback on the documents they are reviewing. If the documents are very similar, lowering the relevancy scale to allow more documents for review could be advisable. This is to ensure there is not confirmation basis and ensure the results are not overly narrow for review.

Know What You are Looking For

The most effective way to respond to discovery requests is to know what you are looking for. This means knowing what data must be found for discovery, claims, or defenses in a lawsuit, in order to have a defensible discovery strategy. If a lawsuit involves discrimination over work schedules, then searches must be created to find that information. At the end of the day, lawsuits are about people. What happened to them? What are the relevant facts? Knowing the answers to those questions can help lawyers develop searches and predictive coding models to identify responsive information.

*See, Mortg. Resolution Servicing v. JPMorgan Chase Bank, N.A., 2017 U.S. Dist. LEXIS 78217, at *6 (S.D.N.Y. May 18, 2017), citing The Sedona Conference, The Sedona Principles: Second Edition, Best Practices Recommendations & Principles for Addressing Electronic Document Production, at ii princ. 6 (2007),