Here is what happened in the recent case of United States v. Mehta, 594 F.3d 277 (4th Cir. 2010):
In establishing the tax loss under § 2T1.1 of the Sentencing Guidelines, the government proposed that the district court consider the 4,321 Schedule A returns filed by Mehta over the four-year period under investigation to find a tax loss of $ 2,508,000. The IRS had selected approximately 941 returns for civil audit by scoring each of the Schedule A returns and choosing those most likely to produce additional tax liability. Of these 941, 775 were selected for a correspondence audit, and the remaining returns were accepted as filed. As part of the correspondence audit, the IRS furnished the taxpayers with a computation of what their additional tax liability would be if they did not produce documentation. Approximately 30% of the taxpayers (or 307 returns) signed IRS Form 4549 agreeing to pay the additional tax assessment. The total additional tax liability for these 307 returns was approximately $ 473,000, and the average tax assessed "per agreed-upon audit" was $ 1,531.The taxpayer first argued that any reliance on returns where the taxpayer signed a Form 4549 was inappropriate because the 4549 is just a waiver of the restrictions on assessment and does not represent the taxpayer’s agreement that the tax is due. The Court rejected that argument in a cryptic analysis that, in my judgment really does not address the point raised:
Of the 4,321 Schedule A returns filed, the district court considered only 2,500 returns which Mehta filed during the last two years of the investigation. The court extrapolated the loss for these 2,500 returns. It multiplied 2,500 by 30% to equal 750 returns, and multiplied 750 by the $ 1,500 average audited tax loss per return to arrive at $ 1,125,000. Therefore, the court calculated a tax loss between $ 1,000,000 and $ 2,500,000 pursuant to Guideline § 2T4.1, resulting in a base offense level of 22. The court applied a two-level increase because Mehta was in the business of preparing tax returns. U.S.S.G. § 2T1.4(b)(1). The Guidelines range for offense level 24 is 51-63 months, but the district court varied downward and sentenced Mehta to 48 months imprisonment.
In addition [to the Forms 4549], an IRS Agent testified at trial that he personally reviewed each of the 307 agreed-upon audited returns and discovered that the fraudulent deductions taken on those returns fit the pattern of fraud evidenced at trial. Based on this evidence, the district court found that the audited returns revealed a "pattern of numbers" reported for various deductions that was strikingly similar to the returns proven fraudulent at trial. Thus, there was ample evidence to support the court's finding that that (sic) it was more probable than not that Mehta fraudulently prepared the audited returns such that they could be used to calculate the tax loss.I fail to see how a Form 4549 alone is any indication that a tax is due. I cannot speculate as to what the IRS Agent may have testified because the Court of Appeals' description is too cryptic, but just for that reason the cryptic description is hardly persuasive that any tax was due for those 307 Form 4549 taxpayers, much less the $1,531 average determined by the Court. And, of course, those 307 taxpayers are the key for the balance of the holding, for the average tax thus determined for those 307 taxpayers was applied to what the district court deemed an appropriate universe in extrapolating the aggregate tax loss. Even here the Court of Appeals gagged, although it ultimately found harmless error. Here is what the Court said:
Second, Mehta argues that the district court erred in multiplying the average tax liability per agreed-upon audit by the 2,500 Schedule A returns filed by Mehta during the last two years of the investigation in calculating the total tax loss. We agree. In conducting that calculation, the court determined that, because 30% of the "flagged for audit" returns had an average tax loss of $ 1,531, this meant that approximately 30% of all the Schedule A returns filed by Mehta during that period would also have an average tax loss of $ 1,531. The problem with this approach is that the 30% figure was derived from a non-random universe of returns that shared the characteristic of having been identified by a computer program as being more likely to contain errors.But, then the Court of Appeals said the error was harmless. The reason was that the district court could have used a larger sample (population in statistics lingo) so that a lower average which was certainly required might have produced a higher tax loss. I think the Court of Appeals' hunch was that properly calculated sampling methodology would have produced a greater tax loss; hence no harm, no foul (or harmless in appellate lingo, meaning the defendant loses regardless of the error). The Court of Appeals' opinion is, in my judgment, hardly persuasive of the conclusion. Hunches are not persuasive evidence upon which to justify a preponderance of the evidence burden. Most of our civil cases are resolved on a preponderance of the evidence standard and many important criminal issues (including most sentencing issues) get resolved on a preponderance of the evidence; I have never seen that standard equated to a reasonable hunch standard.
While extrapolation might, in some cases, be a reasonable method to estimate tax loss, the process used here by the district court goes against the very principles that underlie extrapolation. To extrapolate means "to estimate the values of . . . a function or series . . . outside a range in which some of its values are known, on the assumption that the trends followed inside the range continue outside it." Oxford English Dictionary (2d ed. 1989). As the definition indicates, extrapolation in this case would require a threshold finding that the trend in the known sample, namely the average tax loss in the "flagged for audit" returns, was likely to be present in the larger group of all 2,500 of the Schedule A returns prepared by Mehta during that period. That threshold requirement clearly failed here because the very reason that the "flagged for audit" returns were flagged in the first place was that they were different from the rest of the larger group.
The district court recognized the problem by noting the fact that the "flagged for audit" sample used to calculate the tax loss was not random. The court stated at sentencing, "I just caution for future cases that you can't propose that it's a random sample unless it is." J.A. 1534. The court specifically noted: "We are all speculating that the computer program is designed to flag those most likely to benefit the Government, but I am certainly not in any position to know what those so-called red flags might be, and no one here has been able to enlighten me." J.A. 1529. Because the district court "was certainly not in any position" to understand whether the sample offered was representative of the larger group of 2,500 returns, J.A. 1534, it erred in using that sample to extrapolate the tax loss for the larger group.
I don't doubt that statistical sampling may be an appropriate way to make sentencing tax loss findings. This requires good numbers at each stage -- in picking the appropriate population, in selecting an appropriate random sample from that population that produces an acceptable margin of error, and then applying the results over the larger population (this being the easiest part). That is not what occurred in this case, by the Court of Appeals' own admission.
For a previous blog no this subject, see here.