Commerce Defends Use of Simple Average in 'd' Test Denominator After CAFC Remand

The Commerce Department stuck by its use of a simple average in the denominatory calculation of the Cohen's d coefficient -- a part of the test to root out "masked" dumping -- in Nov. 10 remand results submitted to the Court of International Trade. Responding to an order from the U.S. Court of Appeals for the Federal Circuit telling the agency to justify its departure from the academic literature about how to calculate the Cohen's d denominator, Commerce said that the literature actually supports the use of a simple average when sampling is not used (Mid Continent Steel & Wire v. United States, CIT Consol. #15-00213).

In antidumping duty investigations, such as the present one contested by consolidated plaintiff PT Enterprise on steel nails from Taiwan (see 2107270042), Commerce may identify goods that are dumped into the U.S. market through masked dumping. Since the agency typically conducts its investigations by comparing the average home market price of the good in question with its U.S. price, exporters may work around this by dumping the goods in certain areas and selling them at a higher price in another or at another time. To combat this, Commerce may compare the weighted average of sales in the home country to individual sales prices, called the average-to-transaction comparison method.

Before conducting this analysis though, Commerce must first detect the masked dumping using a differential pricing analysis. The agency breaks down the U.S. sales data into sets based on comparable product groups. Once in the product group, Commerce then breaks that data into various subsets, including the region where the U.S. sales took place, the purchasers involved in the sales and even the time periods of the sales. Commerce will then pick one subset as the "test group" while aggregating the remaining subset into the "comparison group." The agency then employs the Cohen's d test to find whether the test group is significantly different from the comparison group.

The test comes out with a coefficient, or a ratio, wherein the numerator is the difference between the mean prices of the two groups and the denominator is a benchmark from which to judge the significance of this difference. It was the benchmark for the coefficient that served as the point of contention for PT Enterprises' appeal at the Federal Circuit. In April, the appellate court remanded Commerce's position for a second time, finding that the agency's decision to use a simple average of the pooled standard deviation for the Cohen's d denominator instead of a weighted average was unlawful (see 2204210031).

The statistical literature on the d test says that the ideal way to come up with the denominator is to use the standard deviation for the whole population, which can be approximated using a pooled estimate. In the present case, Commerce did not use the entire population's standard deviation, despite acknowledging that it had the whole population's data, opting instead for the simple average. On remand again, Commerce said that, per the court's orders, the agency looked to the academic literature and found that it actually has support for using a simple average when sampling is not used, as is the case here.

This argument from Commerce echoes claims made by the agency in other cases over the use of the Cohen's d test. In other proceedings, Commerce has said that its use of the test does not need to meet certain statistical assumptions, including normal distribution and roughly equal variances, because it is using the entire population of data and not merely a sample (see 2205310059). While the subject of PT Enterprise's appeal is limited more to the use of the simple average in the coefficient's denominator, the argument is largely the same.

"Commerce’s calculation of the Cohen’s d coefficient is not based on sampled data, and there is no estimation of the actual mean and standard deviation of the test group and of the comparison group," the remand results said. "The academic literature provides for the use of a weighted average as a possible approach when estimating the denominator of the effect size when the actual standard deviations are not known, which is not the situation with Commerce’s application of the Cohen’s d test. Therefore, the academic literature allows for the use of the simple average to calculate the denominator of the effect size, and it does not necessarily support the use of a weighted average."

The U.S. further argued that the use of a single standard deviation "is not contemplated by the academic literature," so the agency does not find it to be a reasonable approach for the d test to use a single standard deviation of all the data when it's explicitly split into two groups. "[T]he results of the calculation reflect not just the dispersion of the data within each group, but also the dispersion of the data between the two groups, the precise aspect, i.e., the difference in prices, that the effect size is meant to quantify," Commerce said. "... The academic literature, whether the effect size is based on population parameters or estimates based on statistical samples, bases the calculations on the standard deviations of each group of data, test and comparison group, experimental and control group, and not on an overall, single, commingled group of data."