Commerce Has No Grounds for Use of 'd' Test in Academic Literature, AD Plaintiffs Argue

The Court of International Trade should reject the Commerce Department's continued use of the Cohen's d test following a remand order from the court, plaintiffs, led by Marmen Inc., argued in a July 22 brief at the trade court. Commerce has failed to explain whether limits on the use of the test to detect masked dumping were satisfied with respect to Marmen's data, the brief said. Further, Commerce's premise -- that certain statistical assumptions such as the normal distribution of the data need not apply where the data sets are populations and not samples -- is not supported by anything, including the academic literature, the brief said (Marmen Inc. v. United States, CIT Consol. #20-00169).

The case concerns the antidumping duty investigation on utility scale wind towers from Canada. In the trade court's opinion, Judge Jennifer Choe-Groves upheld four elements of the investigation, but sent back Commerce's decision to reject respondent Marmen's additional cost reconciliation information and use of the average-to-transaction methodology to account for masked dumping (see 2110220069).

On remand, Commerce addressed its differential pricing analysis, continuing to use its average-to-transaction comparison for Marmen, releasing a stouter defense of the Cohen's d test. In July 2021, the U.S. Court of Appeals for the Federal Circuit called the use of the d test into question in Stupp Corp. v. U.S., finding that Commerce did not satisfy certain statistical assumptions, such as normal distribution and roughly equal variances, before running the test (see 2107150032). Choe-Groves remanded Commerce's use of the test in the present case to address the questions raised in Stupp.

On remand, Commerce said that these statistical assumptions are not relevant to its use of the Cohen's d test since it does not use sampled data, which would require these statistical assumptions, but uses the entire data set instead. The agency made the same defense in the Stupp case, which currently sits before the trade court. In response, the plaintiffs in the present case blasted Commerce's rationale.

Support for this position is not found in the academic literature, the brief said. "Critically, Professor Cohen himself used populations -- not samples -- to explain the Coehn’s d coefficient, yet still represented that the test was based on assumptions of normal distributions and equivalent variances," the brief said. Further, the Stupp court cited Robert Grissom and John Kim as greater evidence that the assumptions of normal distributions and equivalent variances must be satisfied to measure the standardized-mean-difference effect size of two populations via the d test.

While Commerce argues that these two authors were describing a similar analysis concerning the overlap of two sets of sampled data, the plaintiffs say this is a false characterization. The authors explicitly reference "populations" in their analysis, but even if Commerce's interpretation is correct, "that does not negate the authors’ representation that 'equal variances' is an assumption that applies to 'populations' -- not only to samples of data," the brief said.

The plaintiffs went on to argue that Commerce failed to address the Federal Circuit's specific concern about the data sets used by Commerce lacking equivalent variances. In Stupp, the Federal Circuit set up a hypothetical wherein the d test produces a large coefficient, though the price differences are practically not significant. Commerce did not address this hypothetical on remand, the plaintiffs said, calling it "ironic" since Commerce justifies its dismissal of the assumptions underlying the test as being a practical move. "This is problematic, because the Appellate Court’s hypothetical -- in which the test group and comparison groups contain 'sales prices that hover around the same value' -- mirrors the issue raised by Marmen in this case," the brief said.