1. Introduction
Edward Leamer opened his 1985 reply to McAleer, Pagan and Volker with one sentence: “A fragile inference is not worth taking seriously.” The reliability factors of the Best Method Rule express the same demand. Treas. Reg. §1.482-1(c)(2)(ii) lists the completeness and accuracy of the data, the reliability of the assumptions, and the sensitivity of the results to possible deficiencies in the data and assumptions. The regulation names a sensitivity test, but transfer pricing practice does not run one. A comparable search applies a sequence of screens in arbitrary order, computes a three-year average ratio profit level indicator (PLI) for each surviving company, and reports the interquartile range (IQR) of the per company average ratios as if no other specification existed.
Leamer advances the discussion, but he is not the last word. His extreme bounds analysis divides the regressors into free variables, which stay in every equation, and doubtful variables, which may enter or leave. The division rests on the judgment of the analyst, and Leamer concedes that extreme bounds analysis “rarely applies” (1985, p. 310). His method measures fragility, but it does not specify the model.
A transfer pricing analysis needs a specification before it needs a sensitivity test. That specification rests on three keys:
- an uncontrolled base in the tested party;
- a reduced-form profit equation, not a structural ratio;
- regression analysis that produces reliable estimates of the comparable PLI with their standard errors.
Leamer’s contribution survives in one place: the report presents a table of alternative estimates, with the assumptions behind each one stated.
Notation follows the Compustat mnemonics: Y is operating profit (OIADP), R is revenue (REVT), X is operating cost (XOPR), D is depreciation (DP), and K is net property, plant and equipment (PPENT). The base of the PLI is written B, where B is R, X or K.
2. First Key: An Uncontrolled Base
The base of the PLI must be free of the controlled transaction in the tested party. Take a distributor t that buys goods from controlled suppliers at transfer price p and sells them to uncontrolled customers. Its operating profit obeys the accounting identity:
\quad (1)\qquad Y_t=R_t-X_t-D_t
Three accounts of this distributor are measured at the transfer price: cost of goods sold (COGS), inventories (INVT), and accounts payable (AP) to the controlled suppliers. Let q_t be the units sold, s_t the units held in stock, and z_t the units purchased but not yet paid for. Abstracting from the inventory costing method, the operating cost divides into an uncontrolled part X_{0t} and a controlled part C_t, and the balance-sheet accounts carry the same price:
\quad
(2)\qquad
X_t=X_{0t}+C_t,\qquad C_t=p\,q_t,\qquad \mathrm{INVT}_t=p\,s_t,\qquad \mathrm{AP}_t=p\,z_tAny variable constructed from these accounts is controlled in the tested party: XOPR, which contains COGS; \Delta\mathrm{INVT}; and any working-capital term built from INVT or AP.
The bases of the comparables are uncontrolled, so the comparables regression is estimated without contamination. The defect appears when the estimated line is applied to the tested party. The benchmark operating profit is the fitted value at the tested party’s base:
\quad
(3)\qquad
\hat{Y}_t=\hat{\lambda}_0+\hat{\lambda}_1 B_tIf the base is XOPR, then B_t = X_t, and by equation (2) the benchmark depends on the price under audit:
\quad
(4)\qquad
\frac{\partial \hat{Y}_t}{\partial p}=\hat{\lambda}_1 q_t>0A higher transfer price enlarges the base and raises the profit the benchmark requires. The benchmark moves with the quantity it is meant to measure, and its standard error is computed at a contaminated point.
If the base is REVT, which is uncontrolled for this distributor, then
\quad
(5)\qquad
\frac{\partial \hat{Y}_t}{\partial p}=0The benchmark is fixed. Equating equation (1) to equation (3) with B_t = R_t yields the arm’s-length price in closed form:
\quad
(6)\qquad
p^{*}=\frac{(1-\hat{\lambda}_1)\,R_t-\hat{\lambda}_0-X_{0t}-D_t}{q_t}The uncontrolled base depends on the position of the tested party in the chain:
| Tested party | Uncontrolled base | Controlled accounts |
|---|---|---|
| Distributor buying from controlled suppliers | REVT | XOPR (contains COGS), INVT, AP |
| Manufacturer selling to an affiliate, inputs bought from uncontrolled parties | XOPR | REVT, receivables (RECT) |
| Either, with equipment bought from uncontrolled parties | PPENT | — |
The regulations treat the choice of base as a question of data availability. The correct criterion is independence from the controlled price.
3. Second Key: A Reduced-Form Profit Equation
3.1 Identity, structure, and reduced form
The choice of Y and B is governed by theory, not by the preference of the analyst. This is the position of Klein and Goldberger (1955), Goldberger (1964), and Kmenta (1986), and it rests on three propositions:
- The accounting identity (1) is exact. It holds by definition and contains nothing to estimate.
- Behavioral (structural) equations are subject to data scouting. Rival structural specifications can always be proposed, and selecting among them by goodness of fit does not settle which is correct.
- The reduced form follows from (1) and (2). It expresses the endogenous variable as a function of the predetermined variables of the system, so it does not require a verdict among rival structural forms.
The reduced form does not abolish the specification problem. It narrows the dispute to the list of predetermined variables, and the identity and the first key settle most of that list.
Within this framework, Leamer’s free variables have no role. No regressor enters because the analyst prefers it. The admissible comparison is between nested models, each tested against the next:
\begin{aligned}
&(7)\quad
\text{M1:}\quad Y_i=\beta B_i+u_i
\\[8pt]
&(8)\quad
\text{M2:}\quad Y_i=\lambda_0+\lambda_1 B_i+v_i
\\[8pt]
&(9)\quad
\text{M3:}\quad Y_i=\lambda_0+\lambda_1 B_i+\lambda_2 K_{i,-1}+w_i
\end{aligned}M1 is the naive univariate model of §1.482-5, which the ratio m_i = Y_i/B_i estimates company by company. M2 is the bivariate model with the uncontrolled base and a free intercept. M3 adds lagged net PPE to test an asset adjustment. Each step tests one restriction, and theory fixes their order: the intercept first, then assets. This answers Leamer’s objection to the procedure of McAleer, Pagan and Volker, in which the order of restrictions was arbitrary (Leamer 1985, p. 312).
3.2 M1 against M2: the ratio as an omitted-variable regression
Dividing M2 by the base gives
\quad
(10)\qquad
m_i=\frac{Y_i}{B_i}=\lambda_1+\lambda_0\left(\frac{1}{B_i}\right)+\frac{v_i}{B_i}The ratio contains a regressor, 1/B, with coefficient \lambda_0. The mean of the ratios is a regression of m on a constant alone, which omits that regressor. The auxiliary regression of 1/B on a constant returns the mean of 1/B, which is 1/H, where H is the harmonic mean of B. Kmenta’s omitted-variable result then gives
\quad
(11)\qquad
E(\bar{m})=\lambda_1+\lambda_0\cdot\frac{1}{N}\sum_{i=1}^{N}\frac{1}{B_i}=\lambda_1+\frac{\lambda_0}{H}The harmonic-mean bias is Kmenta’s formula with the constant as the only included regressor. Along the line, the quantiles carry the same shift by rank inversion: Q_p(m) = \lambda_1 + \lambda_0/Q_{1-p}(B) when \lambda_0 > 0, and \lambda_1 + \lambda_0/Q_p(B) when \lambda_0 < 0. The intercept bias and the specification error are one defect. The ratio imposes \lambda_0 = 0 without testing it, and the bias does not vanish as N increases. Estimating M2 tests the restriction instead of imposing it.
3.3 M2 against M3: relevant and irrelevant variables
Kmenta’s distinction between relevant and irrelevant variables replaces Leamer’s distinction between free and doubtful variables. The two are not equivalent. Kmenta’s distinction is a property of the true model: K_{-1} is irrelevant if \lambda_2 = 0. Leamer’s distinction is a decision of the analyst. Each error in Kmenta’s framework has a known cost. Including an irrelevant K_{-1} leaves \hat{\lambda}_1 unbiased but inflates its variance. Omitting a relevant K_{-1} biases the estimate of \lambda_1:
\begin{aligned}
&(12)\quad
E(\tilde{\lambda}_1)=\lambda_1+\lambda_2\,\delta,
\\[8pt]
&\quad
\delta=\frac{\sum_i (B_i-\bar{B})(K_{i,-1}-\bar{K}_{-1})}{\sum_i (B_i-\bar{B})^2}
\end{aligned}where \delta is the slope of the auxiliary regression of K_{-1} on B. The bias vanishes if \lambda_2 = 0 or \delta = 0.
An insignificant \hat{\lambda}_2 does not prove that K_{-1} is irrelevant, and deleting it whenever a t test fails is pretest estimation. The mean-square-error criterion of Toro-Vizcarrondo and Wallace (1968) weighs the bias of omission against the variance of inclusion. For a single restriction, M2 has a lower mean square error than M3 if and only if
\quad
(13)\qquad
\frac{\lambda_2^{2}}{\operatorname{Var}(\hat{\lambda}_2)}\le 1The criterion is stated in the true parameter, and in practice it is applied with the estimated ratio |\hat{\lambda}_2|/\operatorname{SE}(\hat{\lambda}_2). It agrees with the one-standard-error convention: if \hat{\lambda}_2 \pm \operatorname{SE}(\hat{\lambda}_2) covers zero, M2 is preferred; if the interval excludes zero, K_{-1} stays.
M3 also exposes the conventional asset adjustment. Practitioners adjust the comparables’ margins for differences in asset intensity at an assumed interest rate, which imposes a value on \lambda_2. M3 estimates \lambda_2 and tests it. The imposed adjustment is an untested constrained regression, the same defect as the working capital adjustment.
K_{-1} enters with a lag because year-end PPENT moves with the investment of the same year, so it is not predetermined. It must also satisfy the first key: the tested party must not have bought its equipment from affiliates. For the same reason, \Delta\mathrm{INVT} is not admissible for a distributor that buys from controlled suppliers, because equation (2) measures it at the transfer price.
A side note on the lag: K_{-1} costs one year of data for each comparable. With long series, as in my Springer book, the equation can be estimated one comparable at a time. With the three-year window of a transfer pricing study, the pooled panel can be kept by drawing K_{-1} from the year before the test period.
4. Third Key: Regression Estimates with Standard Errors
The selected model is estimated by least squares on the comparables’ company-years, N = nT for n comparables over T years. With Z the N \times k matrix of regressors and y the vector of operating profits,
\quad
(14)\qquad
\hat{\lambda}=(Z'Z)^{-1}Z'yIn the bivariate model M2, the standard error of the slope is
\quad
(15)\qquad
\operatorname{SE}(\hat{\lambda}_1)=\frac{s}{\sqrt{\sum_{i=1}^{N}(B_i-\bar{B})^2}}where s is the residual standard error. When the residuals are heteroskedastic, or correlated within a company across years, the heteroskedasticity-consistent or company-clustered covariance matrix replaces s^2(Z'Z)^{-1} (Davidson and MacKinnon 2004). The benchmark operating profit at the tested party’s base is equation (3), with standard error
\quad
(16)\qquad
\operatorname{SE}(\hat{Y}_t)=s\,\sqrt{\frac{1}{N}+\frac{(B_t-\bar{B})^2}{\sum_{i=1}^{N}(B_i-\bar{B})^2}}The report states \hat{\lambda}_0 \pm \operatorname{SE}, \hat{\lambda}_1 \pm \operatorname{SE}, and \hat{Y}_t \pm \operatorname{SE}(\hat{Y}_t). Each interval spans one standard error on either side of the estimate, a coverage of about 68 percent. These standard errors fall at the rate 1/\sqrt{N}, so the laws of large numbers favor the analyst: a larger sample yields a more stable estimate with a smaller variance.
4.1 Why the IQR is not a substitute
Treas. Reg. §1.482-1(e)(2)(iii)(B) states that the IQR “ordinarily” provides an acceptable measure of the arm’s-length range, without reference to sample size. The sample quartiles converge to the population quartiles of the PLI, and the distance between those quartiles is fixed: under normality, \mathrm{IQR} \approx 1.349\,\sigma for any N. A larger sample sharpens the estimate of \lambda_1, but it never narrows the IQR. In small samples the quartiles are themselves imprecise. The asymptotic variance of the p-th sample quantile is
\quad
(17)\qquad
\operatorname{Var}(\hat{q}_p)\approx\frac{p(1-p)}{N\,f(q_p)^2}where f is the density of the PLI at its p-th quantile (Kendall and Stuart 1977). With six or eight comparables, each quartile is an interpolation between two observations.
The IQR describes the dispersion of the data around one fixed specification. Equations (15) and (16) describe the uncertainty of the estimate. The regulation treats the first as if it were the second. The same paragraph of the regulation admits a different statistical method if it provides a more reliable measure. Read with the sensitivity factor of §1.482-1(c)(2)(ii), that clause admits regression on the regulations’ own terms.
5. A Table of Alternative Estimates
When readers differ about assumptions, Leamer asks for a menu of inferences, with the assumptions behind each one stated (1985, p. 312). For a transfer pricing report, the menu takes this form:
| Specification | Intercept | Assets | Reported |
|---|---|---|---|
| M1: ratio quartiles (contrast) | Fixed at 0 | None | Q_{25}, Q_{50}, Q_{75} of m |
| M2: OLS | Free | None | \hat{\lambda}_0 \pm \operatorname{SE}, \hat{\lambda}_1 \pm \operatorname{SE} |
| M2: Huber M-estimator | Free | None | \hat{\lambda}_0 \pm \operatorname{SE}, \hat{\lambda}_1 \pm \operatorname{SE} |
| M2: quantile regression 25/50/75 | Free | None | \hat{\lambda}_0 \pm \operatorname{SE}, \hat{\lambda}_1 \pm \operatorname{SE} per quantile |
| M3: OLS | Free | K_{-1} | \hat{\lambda}_2 \pm \operatorname{SE}; shift in \hat{\lambda}_1 from M2 |
All rows use the uncontrolled base. Each row reports the margin implied at the tested party’s own base:
\quad
(18)\qquad
\hat{m}_t=\frac{\hat{Y}_t}{B_t}=\hat{\lambda}_1+\frac{\hat{\lambda}_0}{B_t}\quad\left(+\,\hat{\lambda}_2\,\frac{K_{t,-1}}{B_t}\ \text{in M3}\right)If the rows agree within one standard error, the arm’s-length conclusion is sturdy. If they diverge, the divergence is the finding, and the examiner or the court should see it. In contested cases, the table also demonstrates that the analysis was not selected to favor the assessment or the taxpayer.
6. Conclusion
A reliable PLI rests on three keys: a base free of the controlled price, equations (1) to (6); a reduced-form profit equation, selected by nested tests from M1 to M3, equations (7) to (13); and regression estimates reported with their standard errors, equations (14) to (16). The ratio PLI of §1.482-5 fails the second key by construction. Because a ratio cannot be tested for an intercept, it cannot satisfy the sensitivity factor of §1.482-1(c)(2)(ii). The IQR then reports the dispersion of a misspecified statistic as if it were an arm’s-length range.
Leamer’s diagnostic supplies the form of the report, a table of alternative estimates. It does not supply the specification, and it does not choose the estimator. Those choices come from the accounting identity, the reduced form, and the three keys.
References
Davidson, R., and J. G. MacKinnon. 2004. Econometric Theory and Methods. New York: Oxford University Press. Heteroskedasticity-consistent and clustered covariance matrices.
Goldberger, A. S. 1964. Econometric Theory. New York: Wiley. Structural and reduced forms.
Kendall, M. G., and A. Stuart. 1977. The Advanced Theory of Statistics, Vol. 1, 4th ed. London: Griffin. Sampling variance of order statistics.
Klein, L. R., and A. S. Goldberger. 1955. An Econometric Model of the United States, 1929–1952. Amsterdam: North-Holland. Identities, behavioral equations, and the reduced form.
Kmenta, J. 1986. Elements of Econometrics, 2nd ed. New York: Macmillan. Specification error from omitted relevant and included irrelevant variables.
Leamer, E. E. 1983. “Let’s Take the Con Out of Econometrics.” American Economic Review 73 (1): 31–43.
Leamer, E. E. 1985. “Sensitivity Analyses Would Help.” American Economic Review 75 (3): 308–313. https://www.jstor.org/stable/1814801.
McAleer, M., A. R. Pagan, and P. A. Volker. 1985. “What Will Take the Con Out of Econometrics?” American Economic Review 75 (3): 293–307.
Toro-Vizcarrondo, C., and T. D. Wallace. 1968. “A Test of the Mean Square Error Criterion for Restrictions in Linear Regression.” Journal of the American Statistical Association 63 (322): 558–572.
26 C.F.R. §1.482-1(c)(2)(ii). Best Method Rule: completeness and accuracy of data, reliability of assumptions, sensitivity of results.
26 C.F.R. §1.482-1(e)(2)(iii)(B)–(C). Arm’s-length range; the interquartile range.
26 C.F.R. §1.482-5(b)(4). Profit level indicators.