Two Roads
The road less travelled is Frost’s. This post takes the other.
Deming (1943) and Weisberg (1980) expound one method, least squares, for two problems. Deming adjusts observations: every measured quantity carries error, the regressor included. Weisberg estimates a regression: the mean of one variable given another, with the second free of error.
Deming’s road is the less travelled. Weisberg’s is taken here, and on it one quantity matters for transfer pricing: leverage. Leverage is the input of the HC3 standard error, which is the standard error EdgarStat reports. This post derives HC3 from leverage, one step at a time.
Progressive and Degenerating Expositions
Lakatos (1970) supplies the test. A revision of a research programme is progressive if it predicts what the prior version did not, and the prediction is corroborated. It is degenerating if it accommodates known anomalies by auxiliary devices and predicts nothing new.
The hard core of least squares is fixed since Gauss (1809; 1821–23): minimize the weighted sum of squared residuals. Between Deming and Weisberg, seven elements changed.
| Element | From Deming (1943) to Weisberg (1980; 2014) | Verdict |
|---|---|---|
| Notation | Scalar sums and Lagrange multipliers become matrices | Presentation |
| Computation | Normal equations on a desk calculator become orthogonal decompositions | Presentation |
| Remote observations | Unmeasured becomes leverage h, a number per observation | Progressive |
| Unequal error variances | Weights known in advance become variances estimated from residuals and leverage (HC3) | Progressive |
| Functional form | Assumed becomes tested for lack of fit | Progressive |
| Error in the regressor | Weighted in both coordinates becomes assumed absent | Content lost |
| Equations the true values must satisfy | Imposed on the fit becomes omitted | Content lost |
Two changes are notation. Three add content. Two remove content: regression is the limiting case of Deming’s problem in which X carries no error, and the converse does not hold.
The interquartile range of profit ratios fails the test at the first step. It estimates no parameter. Under normality its width is 1.349 standard deviations at every sample size, while the one-standard-error interval of a mean has width 2\sigma/\sqrt{N}. Its anomalies are met by devices: trimming by rank, working capital adjustments, rejection of comparables without a criterion. That is a degenerating exposition. A result that the opposing party cannot recompute from a stated rule is hearsay.
The Comparables Regression
The regression pools n companies over T years: N = nT company-years. R is revenue (Compustat REVT); X is operating cost (XOPR). The reduced form is
\begin{aligned}
&(1)\quad
R(i,t) = \lambda_0 + \lambda_1 X(i,t) + \varepsilon(i,t),
\\[8pt]
&\quad
i = 1,\dots,n; \quad t = 1,\dots,T .
\end{aligned}\lambda_0 and \lambda_1 are the parameters. \varepsilon(i,t) is the error of company i in year t, with variance \sigma^2(i,t). The errors are taken as independent. Their variances are not taken as equal: a company with ten times the cost of another does not have the same error variance.
Write x(i,t) for the deviation of X from its mean \bar{X}, and S_{xx} for the sum of squared deviations over the N company-years.
\quad
(2)\qquad
x(i,t) = X(i,t) - \bar{X}, \qquad S_{xx} = \sum x(i,t)^2 .Least squares gives the estimate b_1 of the slope \lambda_1, the estimate b_0 of the intercept \lambda_0, and the residuals e(i,t) = R(i,t) - b_0 - b_1 X(i,t).
\quad
(3)\qquad
b_1 = \frac{\sum x(i,t)\, R(i,t)}{S_{xx}} .Step 1: The Variance of the Slope
Substitute the model for R(i,t) in the formula for b_1. The estimate equals the parameter plus a weighted sum of the errors.
\quad
(4)\qquad
b_1 = \lambda_1 + \frac{\sum x(i,t)\, \varepsilon(i,t)}{S_{xx}} .The errors are independent, so the variance of the sum is the sum of the variances.
\quad
(5)\qquad
\operatorname{Var}(b_1) = \frac{\sum x(i,t)^2\, \sigma^2(i,t)}{S_{xx}^2} .This formula is exact. It cannot be computed, because the \sigma^2(i,t) are unknown. Every standard error of b_1 is a rule for replacing them.
Step 2: Three Replacements
The textbook rule sets every \sigma^2(i,t) equal to one value, estimated by s^2. The formula reduces to s^2/S_{xx}. The rule is wrong when the variances differ, which is the ordinary case among companies of different scale.
HC0 (White 1980) replaces each \sigma^2(i,t) by the squared residual e(i,t)^2. No equality of variances is assumed. The defect of HC0 is that a residual is smaller than the error it stands for, and the shortfall depends on leverage.
HC3 (MacKinnon and White 1985) corrects that defect. It requires leverage.
Step 3: Leverage
The leverage of a company-year measures its distance from the mean of X, relative to the dispersion of X in the set.
\quad
(6)\qquad
h(i,t) = \frac{1}{N} + \frac{x(i,t)^2}{S_{xx}} .Leverage lies between 1/N and 1, and the N values sum to 2, the number of coefficients. Its mean is therefore 2/N. Hoaglin and Welsch (1978) call a value above 4/N high.
Leverage governs the size of a residual. If the errors had one variance \sigma^2, the expected squared residual would be
\quad
(7)\qquad
\operatorname{E}\!\left[e(i,t)^2\right] = \sigma^2\left[1 - h(i,t)\right] .A company-year with high leverage draws the fitted line toward itself. Its residual is small for that reason, not because its error is small. HC0 therefore understates the variance most for the companies that determine the slope.
Step 4: The Deleted Residual
Fit the regression without company-year (i,t), and predict its revenue from that fit. The deleted residual d(i,t) is the observed revenue less that prediction. It is the error of the line for a company-year the line has not seen. No second regression is needed, because of an exact identity:
\quad
(8)\qquad
d(i,t) = \frac{e(i,t)}{1 - h(i,t)} .The deleted residual is the ordinary residual divided by one minus the leverage. Where leverage is high, the division restores what the fit removed.
Step 5: HC3
HC3 replaces each \sigma^2(i,t) by the squared deleted residual.
\begin{aligned}
&(9)\quad
V_3(b_1) = \frac{\sum x(i,t)^2\, d(i,t)^2}{S_{xx}^2}
\\[8pt]
&\phantom{(9)\quad V_3(b_1)}
= \frac{1}{S_{xx}^2} \sum \frac{x(i,t)^2\, e(i,t)^2}{\left[1 - h(i,t)\right]^2} .
\end{aligned}The divisor is squared because the deleted residual is squared. That answers the recurrent question of why HC3 divides by (1-h)^2 and not by (1-h). The standard error is the square root: SE_3(b_1) = \sqrt{V_3(b_1)}.
In matrix form, for both coefficients, with \mathbf{X} the N \times 2 matrix whose first column is a column of ones, and \mathbf{S} the diagonal matrix of the d(i,t)^2:
\quad
(10)\qquad
\mathbf{V}_3 = (\mathbf{X}'\mathbf{X})^{-1}\, \mathbf{X}'\mathbf{S}\mathbf{X}\, (\mathbf{X}'\mathbf{X})^{-1} .The table shows the factor by which HC3 enlarges a squared residual, for n = 8 companies and T = 3 years, N = 24.
| Company-year | Leverage h | Factor 1/(1 − h)² |
|---|---|---|
| At the mean of X | 1/N = 0.042 | 1.09 |
| Average leverage | 2/N = 0.083 | 1.19 |
| High leverage (Hoaglin and Welsch) | 4/N = 0.167 | 1.44 |
| One company dominates the set | 0.500 | 4.00 |
What HC3 Does for Transfer Pricing
The interquartile range removes companies. HC3 removes none. Each company-year stays in the regression, and its leverage enters the standard error of the slope. A set dominated by one large company yields a wider interval, by a stated formula.
Weisberg also teaches Cook’s distance, a measure for deciding which observations to delete. Deletion is a decision, and a hostile examiner can contest it company by company. HC3 requires no such decision. That is the reason EdgarStat takes HC3 and not deletion.
The report is the slope with one standard error: b_1 \pm SE_3(b_1). Under the model this interval covers \lambda_1 with probability near 68%. The intercept is reported the same way: b_0 \pm SE_3(b_0).
The Protocol
- Estimate the regression of R on X over all N company-years. Obtain b_0, b_1, and the residuals e(i,t).
- Compute the leverage h(i,t) of each company-year.
- Compute the deleted residuals d(i,t) = e(i,t)/[1 - h(i,t)] and the HC3 standard errors.
- Report b_0 and b_1, each with one HC3 standard error.
- Publish h(i,t) and the factor 1/[1 - h(i,t)]^2 for each company-year, so that the reader sees which companies widen the interval.
- Remove no company by rank.
Treas. Reg. § 1.482-1(e)(2)(iii)(B) accepts the interquartile range in the ordinary case and permits a different statistical method that provides a more reliable measure. Every number in the protocol follows from a formula in this post, and the opposing expert can recompute each one.
What This Road Omits
HC3 corrects for unequal error variances. It does not correct for two other conditions.
First, HC3 takes the N company-years as independent. The T years of one company are not independent, and HC3 does not address that dependence.
Second, leverage and HC3 take X as exact. Reported operating cost carries error, and neither quantity detects it. Deming’s adjustment treats error in X; regression does not. That is the content lost between 1943 and 1980, and the subject of a later post.
Summary Conclusion
Leverage h(i,t) measures how far a company-year lies from the mean of X. A residual understates the error by the factor 1 - h(i,t). HC3 restores the shortfall by dividing each residual by 1 - h(i,t) before squaring it. Deming (1943) had no such correction; the regression road supplies it.
The interquartile range selects companies by rank and estimates no parameter. A comparables regression with HC3 standard errors retains every company, estimates the slope, and states its uncertainty by a formula that both parties can repeat.
References
Aitken, A. C., and Silverstone, H. (1942). On the estimation of statistical parameters. Proceedings of the Royal Society of Edinburgh, 61, 186–194. Cited by Deming; minimum-variance estimation.
Davidson, R., and MacKinnon, J. G. (2004). Econometric Theory and Methods. Oxford University Press. Textbook treatment of leverage and of HC0 to HC3.
Deming, W. E. (1943). Statistical Adjustment of Data. Wiley. Dover reprint, 1964. Least squares with error in all observed quantities.
Frost, R. (1916). The Road Not Taken. In Mountain Interval. Henry Holt.
Gauss, C. F. (1809). Theoria motus corporum coelestium. Perthes and Besser. The first justification, from the normal law.
Gauss, C. F. (1821–23). Theoria combinationis observationum erroribus minimis obnoxiae. English translation by G. W. Stewart, SIAM, 1995. The second justification, without the normal law.
Hoaglin, D. C., and Welsch, R. E. (1978). The hat matrix in regression and ANOVA. The American Statistician, 32, 17–22. Leverage and the 2p/N rule.
Lakatos, I. (1970). Falsification and the methodology of scientific research programmes. In I. Lakatos and A. Musgrave (eds.), Criticism and the Growth of Knowledge. Cambridge University Press. Progressive and degenerating problem shifts.
MacKinnon, J. G., and White, H. (1985). Some heteroskedasticity-consistent covariance matrix estimators with improved finite sample properties. Journal of Econometrics, 29, 305–325. Source of HC3.
Weisberg, S. (1980). Applied Linear Regression. Wiley. Fourth edition, 2014. Leverage, residuals, and deletion in a first course on regression.
White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity. Econometrica, 48, 817–838. Source of HC0.
26 C.F.R. § 1.482-1(e)(2)(iii)(B). The interquartile range and the allowance for a more reliable statistical method.