Abstract: This article demonstrates algebraically that profit ratios computed as m(i) = Y(i)/X(i) are biased estimators of the true economic slope parameter when the underlying relationship Y = a + bX has a non-zero intercept. The bias equals a/H, where H is the harmonic mean of X. Consequently, the quartiles of ratio-based profit level indicators (PLIs)—the standard method in transfer pricing practice—produce distorted (unreliable) arm’s length ranges. Regression analysis provides an unbiased and reliable alternative.
1. The Problem: Ratios as Slope Estimators
Transfer pricing practitioners compute profit level indicators (PLIs) as simple ratios of financial variables—for example, operating profit divided by sales, assets, or costs. These ratios are then arrayed, and the interquartile range is offered as the “arm’s length range.” This procedure treats the ratio m(i) = Y(i)/X(i) as if it were an unbiased estimator of the relationship between Y and X. This hypothesis must be tested using comparables to the “tested party” on a case-by-case basis.
The critical assumption embedded in ratio analysis is that the functional relationship between the numerator Y and denominator X passes through the origin—that is, the relationship is strictly proportional. When this assumption fails, the ratio becomes a biased estimator, and the magnitude of the bias depends on the distribution of X itself.
2. The Linear Regression Model
Assume that the true relationship between Y(i) and X(i) for a group of comparable enterprises i = 1, 2, …, N is given by:
\begin{aligned}
Y(i) = a + bX(i) \\
\text{Equation (1)}
\end{aligned}where a is the intercept, b is the slope coefficient, and X(i) > 0 for all i. No stochastic error term is necessary to demonstrate the bias; the result is algebraic, not statistical.
In transfer pricing, Y represents operating profit, and X represents operating assets, sales revenue, or total costs. The slope b represents the marginal profit rate—the additional profit associated with an additional unit of X. This is the economically meaningful parameter.
3. The Ratio as an Estimator
Practitioners compute the profit ratio for each enterprise:
\begin{aligned}
\quad m(i) &= \frac{Y(i)}{X(i)} \\[6pt]
\end{aligned}Substituting Equation (1) into this expression:
\begin{aligned}
\quad m(i) &= \frac{a + bX(i)}{X(i)} \\[6pt]
\end{aligned}Distributing the division:
\begin{aligned}
\quad m(i) &= \frac{a}{X(i)} + b \\
&\text{Equation (2)} \\[10pt]
\end{aligned}Equation (2) is the central result. Each computed ratio m(i) equals the true slope b plus a term that depends on the intercept and the reciprocal of X(i). When a ≠ 0, the ratio m(i) is not equal to b—it is contaminated by the bias term a/X(i).
4. The Mean of the Ratios and the Harmonic Mean
To determine the expected value (or arithmetic mean) of the ratios across all N enterprises, compute:
\begin{aligned}
\text{Mean}[m(i)] &= \frac{1}{N} \sum m(i) \\
&= \frac{1}{N} \sum \left( \frac{a}{X(i)} + b \right) \\
&= \frac{1}{N} \sum \frac{a}{X(i)} + \frac{1}{N} \sum b \\
&= a \cdot \frac{1}{N} \sum \left( \frac{1}{X(i)} \right) + b
\end{aligned}Recall that the harmonic mean H of X(i) is defined as:
\begin{aligned}
\quad H &= \frac{N}{\sum \left(\frac{1}{X(i)}\right)} \\[6pt]
\end{aligned}Therefore:
\begin{aligned}
\quad \frac{1}{N} \sum \left(\frac{1}{X(i)}\right) &= \frac{1}{H} \\[10pt]
\end{aligned}Substituting:
\begin{gathered}
\text{Mean}[m(i)] = b + \frac{a}{H} \\
\text{Equation (3) — The Harmonic Mean Bias Theorem}
\end{gathered}The arithmetic mean of the ratios m(i) equals the true slope b plus a bias term a/H. This bias vanishes if and only if a = 0—that is, ratio analysis is reliable if and only if the true relationship passes through the origin.
5. Direction and Magnitude of the Bias
The bias a/H has important properties:
Sign: The bias has the same sign as the intercept a. A positive intercept (fixed costs that must be covered regardless of scale) inflates the mean ratio above b. A negative intercept (economies of scale in profitability) deflates the mean ratio below b.
Magnitude: The bias is inversely proportional to the harmonic mean H of the denominators X(i). Since the harmonic mean is sensitive to small values—H is always less than or equal to the arithmetic mean, with equality only when all X(i) are identical—samples containing small enterprises will exhibit larger bias.
Heterogeneity: The bias is not uniform across enterprises. From Equation (2), enterprise i carries bias a/X(i). Small enterprises (small X) are more heavily biased than large enterprises. This heteroskedasticity in bias propagates to the quartiles.
6. Bias in the Quartiles
Transfer pricing practice does not use ratios; it uses their quartiles. Does the bias affect quartiles? Consider the first quartile Q₁, defined as the value below which 25% of the ratios fall.
From Equation (2):
\begin{aligned}
\quad m(i) &= b + \frac{a}{X(i)} \\[6pt]
\end{aligned}The ranking of m(i) across enterprises depends on the values of a/X(i). If a > 0, enterprises with larger X(i) have smaller ratios (closer to the true slope b), while enterprises with smaller X(i) have inflated ratios. The quartiles inherit this distortion.
If we order enterprises by X(i) and denote the corresponding order statistic of X as X₍₁₎ < X₍₂₎ < … < X₍ₙ₎, then when a > 0:
\begin{aligned}
\quad m_{(1)} &> m_{(2)} > \dots > m_{(n)}
\end{aligned}The smallest enterprises generate the largest ratios. The first quartile Q₁ of ratios corresponds to large enterprises (low bias), while the third quartile Q₃ corresponds to small enterprises (high bias). The interquartile range captures neither a reliable distribution of profitability nor a symmetric band around the true slope.
7. The Untested Assumption About the Intercept
The ratio method implicitly assumes that the intercept a = 0. This assumption is not tested in standard transfer pricing practice. Practitioners compute ratios, sort them, extract quartiles, and call it a day. At no point does the methodology ask: “Is the underlying relationship actually proportional?”
This is not a minor technical detail. The assumption that a = 0 is a strong functional form restriction—equivalent to assuming that an enterprise with zero assets, zero sales, or zero costs would have exactly zero profit. Economically, this is implausible. Enterprises have fixed costs, threshold profitability requirements, and economies of scale that create non-zero intercepts.
The ratio method imposes a stronger linearity assumption than regression. Regression estimates both a and b, allowing the data to determine whether a = 0. The ratio method assumes a = 0 without verification, forcing all variation into the slope. When this assumption is false, the quartiles are biased.
8. The Regression Analysis Remedy
Ordinary least squares (OLS) regression of Y on X produces unbiased estimates of both a and b under the Gauss-Markov conditions. The estimated slope b̂ is the best linear unbiased estimator (BLUE) of the marginal profit rate. Otherwise, use generalized least squares (GLS); both algorithms are available as a service in EdgarStat.
Also, regression provides diagnostics, including the t-statistic for the intercept. If the null hypothesis H₀: a = 0 is rejected, the ratio method is demonstrably inappropriate for that data. If the null is not rejected, proportionality may be acceptable—but this is an empirical finding, not an assumption. The R-squared statistic assesses the strength of the relationship between X and Y. None of these advantages is available when using quartiles of ratios, such as profit indicators.
Regression also produces:
Fitted values: Ŷ(i) = â + b̂ X(i), which serve as the arm’s length benchmark for each tested party given its actual X.
Confidence intervals: Confidence bands that account for both estimation uncertainty and residual variance, providing a statistically grounded arm’s length range of the selected profit indicator.
Heteroskedasticity diagnostics: Tests (Breusch-Pagan, White) that detect whether the variance of the error term depends on X, indicating when generalized least squares (GLS) should replace OLS. Instead of two-phase testing, we prefer using GLS or robust regression.
9. Implications for Transfer Pricing
The consequences for arm’s length analysis are direct:
1. The interquartile range is not centered on the true parameter. When a > 0, all quartiles are inflated. When a < 0, all quartiles are deflated. The “arm’s length range” systematically misses the arm’s length result if the intercept of the linear relationship is nonzero.
2. Adjustments to the median or quartile compound the error. Treasury Regulations permit adjustment to the median when the tested party’s result falls outside the interquartile range. If the range itself is biased, the adjustment “corrects” to a wrong benchmark.
3. Size-based stratification does not eliminate bias. Practitioners sometimes filter comparables by size. This changes H but does not make a = 0. The bias persists, merely rescaled.
4. The “most reliable measure” standard is violated. 26 CFR §1.482-1(e) requires that the arm’s length result be determined using the method that provides the most reliable measure. A biased estimator is, by construction, less reliable than an unbiased estimator. Ratio analysis cannot satisfy this standard when the intercept is nonzero.
10. Conclusion
The quartiles of profit ratios are not delivered to practitioners by divine intervention. Quartiles of profit indicators are not deus ex machina. They are computed from data under an implicit and untested assumption—that the intercept of the underlying profit function is zero. When this assumption is false, and economic reasoning and empirics suggest it usually is, the resulting arm’s length range is biased by a/H, where H is the harmonic mean of the denominators.
Regression analysis dispenses with this assumption. It estimates the intercept, tests whether proportionality holds, and produces unbiased estimates of the economically relevant parameter—the marginal profit rate. The fitted values and confidence intervals from regression analysis provide arm’s length benchmarks that are both theoretically grounded and empirically testable.
The ratio method has persisted in transfer pricing practice for reasons unrelated to its statistical properties. It is computationally simple, requires no econometric training, and produces a fashionable result. That the result is biased has not troubled practitioners. Perhaps it should, because unreliable profit indicators are prone to defeat under audit scrutiny or cross-examination.
References
26 CFR §1.482-1(e), Best Method Rule.
Aitken, A.C. (1936). On Least Squares and Linear Combinations of Observations. Proceedings of the Royal Society of Edinburgh, 55, 42–48.
OECD (2022). Transfer Pricing Guidelines for Multinational Enterprises and Tax Administrations. Paris: OECD Publishing.