Sample. The opposing report, data and results are invented and describe no real matter or expert. This sample is written from the defense side to show that the analysis is the same whichever side retains it.

Rebuttal Analytics Memorandum

Review of the plaintiffs' statistical report on promotion rates — what survives and what does not

To	[Defense counsel], [Firm]
From	Karim Souidi, M.Sc. (Econometrics), NorthLaw.ai
Matter	[Plaintiffs] v. [Employer] — sample file NL-2026-0002 (Title VII, promotion; putative class)
Date	25 September 2026
Purpose	Consulting review of the report of [Opposing Expert] dated [date]. Privileged; prepared at the request of counsel. Not prepared for filing.

1. Bottom line

The opposing report concludes that women were promoted at a rate 2.9 standard deviations below expectation company-wide and that this disparity is "highly statistically significant". Three of the report's choices drive that result, and each is contestable: it counts every employee as eligible for promotion rather than those who applied or met the posted criteria; it pools four job families with different promotion ladders into one test; and it does not adjust for the fact that the promotion decisions were made in 27 separate departments.

When eligibility is defined by the employer's own posted criteria and the analysis respects department structure, the company-wide disparity falls to 1.4 standard deviations (adjusted odds ratio 1.6, 95% CI 0.9 – 2.7), which is not statistically significant by the standard the report itself invokes. One job family, Field Operations, retains a significant disparity on its own (2.4 SD); the other three do not. Counsel should expect the plaintiffs to narrow their theory to that family, and the exposure discussion should start there.

2. Question presented

Whether the opposing report's statistical conclusions are supported by its data and methods; which findings survive correction of identifiable errors; and what the corrected analysis implies for class-wide and job-family-specific claims.

3. What the opposing report did

Using the employer's HRIS extract (4,860 employees, 2021 – 2024), the report compared the share of promotions received by women (31.2%) with women's share of headcount (38.9%), computed a standard-deviation statistic on that comparison, and supplemented it with a logistic regression of promotion on gender with tenure and rating as controls. It reports 2.9 SD and an odds ratio of 0.66 (p = 0.003).

4. Findings

Issue	What the report did	Why it matters	Effect on SD
1. Eligibility pool	Treats all employees as eligible for promotion each year	The employer's postings require a minimum tenure and rating; 41% of headcount is ineligible in a given year, and ineligibility differs by gender (more recent female hires)	2.9 → 2.1
2. Aggregation	Pools four job families into one company-wide test	Promotion rates and gender mix differ sharply by family; pooling creates a Simpson's-paradox pattern (Appendix A2)	2.1 → 1.7
3. Decision unit	Standard errors assume independent individual decisions	Decisions were made within 27 departments; clustering standard errors by department widens the interval	1.7 → 1.4
4. Double counting	Lateral transfers with a title change coded as promotions	112 transfers (63% male) counted as promotions; removing them slightly increases the disparity	1.4 → 1.5
5. Selective reporting	Reports the one subgroup (Field Operations) with a significant result, not the three without	Not an error in itself, but the report's text implies the finding is company-wide	—

Effects are cumulative in the order shown; the sequence matters little (Appendix A3).

5. What survives

Field Operations (1,140 employees, 22% women): a promotion disparity of 2.4 SD with eligibility applied and clustering respected, OR 0.52 (95% CI 0.31 – 0.87). This finding is robust and counsel should treat it as real.
Company-wide: no significant disparity once the three corrections are made. The report's headline number is a consequence of methodological choices, not of the data.
Commonality: the fact that the disparity is confined to one job family with its own promotion ladder is itself evidence against a common company-wide policy, which bears on class certification.

6. Limits of this review

This review uses the same HRIS extract as the opposing report; neither side has verified it independently.
Eligibility was defined from the employer's posted criteria. If plaintiffs show the criteria were applied inconsistently, the eligibility correction weakens.
A non-significant company-wide result does not establish the absence of discrimination; it establishes that the data do not demonstrate it by the standard the opposing report relies on.
Prepared for counsel's use; a filed rebuttal report under FRCP 26(a)(2)(B) would include the full replication log and code.

Karim SouidiM.Sc. Econometrics · DASCA Senior Data Scientist · NorthLaw.ai · karim@northlaw.ai

Appendix A — Technical results

A1. Replication. The opposing report's headline figures were reproduced exactly from the produced extract (2.9 SD; OR 0.66, p = 0.003), confirming that the differences below arise from method, not data.

Table A2. Promotion rates by job family, eligible employees only, 2021 – 2024.

Job family	Eligible (F / M)	Promoted (F / M)	Rate F	Rate M	SD (clustered)	OR (95% CI)
Field Operations	168 / 612	19 / 121	11.3%	19.8%	2.4	0.52 (0.31 – 0.87)
Corporate	402 / 388	71 / 66	17.7%	17.0%	0.3	1.05 (0.72 – 1.53)
Sales	221 / 347	40 / 68	18.1%	19.6%	0.4	0.91 (0.58 – 1.41)
Technical	184 / 545	33 / 104	17.9%	19.1%	0.3	0.93 (0.60 – 1.44)
All families, pooled (opposing method)	1,890 / 2,970	—	—	—	2.9	0.66 (0.50 – 0.87)
All families, eligible, stratified, clustered	975 / 1,892	163 / 359	16.7%	19.0%	1.4	0.84 (0.66 – 1.08)

A3. Order of corrections. Applying the three corrections in all six possible orders yields final values between 1.3 and 1.5 SD; the conclusion does not depend on sequence.

A4. Method. Eligibility defined per posting as ≥ 24 months tenure and rating ≥ 3 in the prior cycle. Stratified (Mantel-Haenszel) comparison across job family × year; logistic regression of promotion on gender with tenure, rating, job family, year and department fixed effects; standard errors clustered by department (27 clusters; wild-cluster bootstrap gives equivalent intervals). Transfers identified from the HRIS action code and excluded.

Appendix B — Data

Source: employer HRIS extract as produced to plaintiffs (Bates [range]), identical to that used by the opposing expert. Population: 4,860 employees with at least one active year 2021 – 2024; 2,867 employee-years eligible under posted criteria. Fields: gender, hire date, job family, department, annual rating, action codes (promotion, transfer, separation). No records excluded beyond ineligible employee-years and coded transfers. Data not independently verified.