concurring in the judgment:
Having written Parts I and H.A., I write separately on Part II.B. because I cannot join the majority in finding clear error (albeit, harmless) in the district court’s calculation of the tax loss. The Guidelines recognize that “when indirect methods of proof are used, the amount of tax loss may be uncertain”; therefore, a district court is not required to find a precise figure but is instead required to make a “reasonable estimate based on available facts.” U.S.S.G. § 2T1.1, cmt. n. I.1 In making a reasonable estimate of the total amount of tax loss where multiple instances of fraud are alleged, a sentencing court may extrapolate the average amount of loss from a sample of audited returns and apply that average to the remaining returns for which the amount of loss is unknown. See United States v. Bryant, 128 F.3d 74, 76 (2d Cir.1997) (finding the district court’s estimation of total loss based on an average loss of less than $100 per unaudited return where audits revealed an average loss of more than $2,400 per return to be “highly generous”). There is no requirement that the sample be random so long as the district court takes measures to make sure that the estimate is a reasonable one.
The district court’s calculation of tax loss here was a reasonable estimate. Because the district court questioned whether the sample of audited returns was representative of the large group, it took measures to compensate for any error in the extrapolation by drastically reducing the universe of returns. The court reduced the number of Schedule A returns it considered from 4,321 to 2,500 by taking only two years of returns despite its finding that Mehta falsified returns for at least four tax years. It then considered only 30% (750) of that 2,500. Then, considering only 750 returns, the court extrapolated the average tax loss of the audited returns to arrive at an amount that exceeds $1 million dollars. Viewing the conservative *285methodology employed by the district court in its loss calculation, I conclude that the amount of tax loss was a reasonable estimate based on available facts and that the district court did not clearly err.2
. Cf. United States v. Pierce, 409 F.3d 228, 234 (4th Cir.2005) (upholding district court's estimation of loss caused by fraudulent scheme based on witness's testimony regarding the number of purchases made and that the conspiracy maintained the same level of purchasing activity throughout the duration of the conspiracy).
. Reasonableness is necessarily a case-by-case determination, and in some instances, a nonrandom sample cannot be used to make a reasonable estimate. If, based on the way the audited returns were selected (for example, based on the amount of income reported), the average tax loss per audited return is likely to be much higher than the tax loss for the unaudited returns, the district court may not be able to compensate for the skewed sample by merely reducing the universe of returns as it did here or by reducing the average tax loss used to extrapolate as was done in Bryant. See, e.g., United States v. Ahanmisi, 324 Fed.Appx. 258 (4th Cir.2009).