Commerce Applied Cohen's 'd' Test Despite Unsuitable Data, Steel Exporter Argues

The Commerce Department's continued use of the Cohen's d test on remand in an antidumping duty case included data that didn't meet the test's statistical criteria, exporter SeAH Steel said in its Aug. 9 remand comments at the Court of International Trade. Commerce cherry-picked data samples out of context to support its arguments and then failed to provide an explanation of how the department's use of the d test showed a pattern of SeAH’s U.S. prices that differed significantly among purchasers, regions or periods, SeAH said. SeAH asked the court to again remand the case to Commerce for reconsideration, especially regarding the limitations on the "reasonable use of Cohen’s d test" (Nexteel Co. v. U.S., CIT Consol. # 18-00083).

In his own writing, statistician Jacob Cohen, the test's creator, "stated explicitly" that his proposed thresholds for evaluating the d statistic were intended to be used only with a “t-test” of statistical significance. Commerce admitted in other cases that such a t-test requires that the assumptions of normality, equal variances and sufficient data be used. SeAH said that its pricing data "did not follow a Normal distribution, and it did not have equal variances or a sufficient number of data points in the groups being compared," making the use of the test inappropriate. "Commerce's conclusions have no support in the academic materials, and are in fact contradicted by Professor Cohen himself," SeAH said.

The redetermination makes no effort to address a "fundamental mathematical issue," and instead relied on an argument that Cohen’s proposed thresholds have universal validity because of their "wide acceptance." In its remand results, Commerce said that its selection of the 0.8 threshold was set by Cohen and that the threshold had been "widely accepted" within the academic community (see 2306280030).

The "real-world examples" referenced by Commerce as evidence for the test's applicability all came from studies of normally distributed data, SeAH said. The issue now isn't whether "Cohen’s proposed thresholds were supported by real-world phenomena" but "whether the phenomena on which Professor Cohen purported to rely provide an appropriate yardstick for measuring other, unrelated phenomena," SeAH said.

Commerce’s attempt to cherry-pick examples out of the context in which Cohen used them is "academically dishonest," SeAH said. The agency needed to show a connection between SeAH’s prices and the example data sets considered by Cohen, but it didn't, the company argued.

SeAH's case is one of many working their way through the courts, challenging the use of this test, and many center on the U.S. Court of Appeals for the Federal Circuit's ruling in Stupp Corp. v. U.S., in which the appellate court questioned the use of the test. The case has now returned to the Federal Circuit after CIT validated Commerce's explanation that the statistical assumptions need not apply given the use of a whole population of data instead of a sample (see 2307260057).