Ordinary least squares (OLS) produce two things: an estimate of the intercept and slope coefficients, and an estimate of their uncertainties. The coefficient estimates are sound, based on matrix algebra inversion. The standard errors rest on an assumption that is seldom true of benchmark comparables — that every comparable’s deviation from the fitted line is drawn from a distribution of the same width. In practice, larger comparables deviate by larger absolute amounts. When that assumption fails, the reported standard error is unreliable, and with it the interval placed around the estimate.
A heteroskedasticity-consistent (HC) standard error estimator (such as HC3) repairs the standard error without touching the coefficient estimate. The point estimate is unchanged; only the stated uncertainty is corrected. The HC3 standard error correction of the slope coefficient takes the form:
\quad \operatorname{Var}(\hat{\beta})
=
(X'X)^{-1}M(X'X)^{-1}\\
M
=
\sum_i
\left[
\frac{e_i^2}{(1-h_i)^2}
\right]
X_iX_i'where e_i is the residual for comparable i and h_i (non-diagonal element of hat matrix, a number between zero and one) is the individual leverage — the extent to which that observation, by virtue of its position on the X-axis, pulls the fitted line toward itself. M is the middle matrix of the sandwich.
Leverage Adjustment
White’s original estimator (HC0), HC3 transforms squared residuals by dividing them by (1 − h_{ii})^2, where h_{ii} represents the leverage (hat matrix diagonal element) of observation i.
Why HC3 rather than HC0, HC1, or HC2
Four variants of the middle matrix are in circulation. HC0 uses {e_i}^2 unmodified. HC1 applies a degrees-of-freedom scale. HC2 divides by (1 − h_i). HC3 divides by (1 − h_{ii})^2. Among the four it is the widest in almost every sample, which makes it the conservative choice at the sample sizes prevailing in transfer pricing: benchmark sets of n = 5 to 15 comparables, so nT between 15 and 45. At those sizes, the unmodified residuals are systematically too small at high-leverage points, and HC0 and HC1 understate uncertainty as a result. HC3 corrects for this and errs toward the wider interval.
The practical consequence for a corporate tested party is that the reported interval is not narrowed by an untested assumption about the dispersion of the comparables. The estimator is a documented, standard choice, not an arbitrary one, and it is the conservative choice among the available alternatives — which is the relevant property when the analysis is examined by an adverse party.
Application within EdgarStat
The estimator is applied to the operating profit markup equation:
\mathrm{REVT}
=
\lambda_0+\lambda_1\mathrm{XOPR},
\qquad
\mathrm{XOPR}
=
\mathrm{COGS}+\mathrm{XSGA}estimated across the benchmark set, each comparable contributing three years of reported data. \lambda_1 is the markup factor on total operating expense; \lambda_0 is the intercept that ratio-based profit indicators assume to be zero without testing it. The estimated markup is reported as one-standard-error interval:
\hat{\lambda}_1
\pm
SE(\hat{\lambda}_1)and it is that standard error (SE) — not the coefficient — that HC3 corrects.
The correction is not decorative in this regression specification of the selected PLI. Unequal variance is structural here rather than incidental: a comparable with fifty million dollars of operating expense and one with five billion cannot produce residuals of similar magnitude, and the large comparable carries high leverage, sitting far out on the X-axis where it exerts disproportionate pull on the fitted line. These are the observations whose residuals ordinary least squares understate most severely, and the division by (1 − h_i)^2 is aimed at correcting them. Using conventional (OLS) standard errors on this equation would report a confidence interval narrowed by an assumption the data contradict — a defect an examining agent is entitled to find, and one that is avoidable at no cost to the point estimate.
Essential references
White, H. (1980). “A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity.” Econometrica 48(4), 817–838. The source of the heteroskedasticity-consistent sandwich estimator in econometrics — HC0, the unadjusted form from which HC1 through HC3 are derived. Establishes that valid standard errors are obtainable without specifying the form of the unequal variances, only their presence. This is why the estimator survives cross-examination: it does not require the analyst to model something unobservable. The underlying result has antecedents in Eicker (1963) and Huber (1967), and the estimator is sometimes cited under all three names.
MacKinnon, J. G., and White, H. (1985). “Some Heteroskedasticity-Consistent Covariance Matrix Estimators with Improved Finite Sample Properties.” Journal of Econometrics 29(3), 305-325. Introduces HC2, after Horn, Horn, and Duncan (1975), and the jackknife estimator, and shows by simulation that White’s original estimator is unreliable in small samples. HC1 is the degrees-of-freedom correction of Hinkley (1977). The HC3 formula reported here, with division by (1 - h_{ii})^2, is the approximation given in Davidson, R., and MacKinnon, J. G. (1993), Estimation and Inference in Econometrics, New York: Oxford University Press.
Long, J. S., and Ervin, L. H. (2000). “Using Heteroscedasticity Consistent Standard Errors in the Linear Regression Model.” The American Statistician 54(3), 217–224. The applied recommendation: HC3 should be the default whenever N is below roughly 250. Transfer pricing benchmark sets sit an order of magnitude below that threshold, which places this specification squarely inside the recommendation rather than at its margin.
Greene, W. H. (2018). Econometric Analysis, 8th ed. New York: Pearson. Chapter on the classical linear regression model, section on robust estimation of the covariance matrix. Included to establish that HC3 standard error is textbook material rather than a specialist construction. Section and page vary by edition and should be verified against the edition cited.
Hayes, A.F., Cai, L. “Using heteroskedasticity-consistent standard error estimators in OLS regression: An introduction and software implementation.” Behavior Research Methods 39, 709–722 (2007). https://doi.org/10.3758/BF03192961