Leverage and Influence

November 19, 2026 by Ednaldo Silva

Introduction

A comparables regression of revenue on operating cost is usually estimated on few observations; for example, eight companies over three years give N=24 company-years. In samples of this size one large company can influence the slope, and the analyst must decide whether that company informs the estimate or distorts it. Two distinct properties are at stake. Leverage is the distance of a company-year from the center of the sample in operating cost (independent variable). Influence is the effect of that company-year on the estimate. A distant observation that lies on the line is an asset: it extends the range of the explanatory variable and sharpens the slope. A distant observation that lies off the line is a liability. This note derives both measures, and states the boundaries EdgarStat applies.

1. The model

For company i in year t,

\quad
(1)\qquad
R(i,t) = a + b\,X(i,t) + e(i,t),

where R is revenue and X is operating cost. The slope b is one plus the markup on total cost; the intercept a absorbs fixed elements of cost and revenue. Estimation is by ordinary least squares, with p=2 parameters and N company-years.

2. The hat matrix

Write the regressors as the N\times 2 matrix Z=[\mathbf{1},X]. The fitted revenues are

\quad
(2)\qquad
\hat R = Z(Z'Z)^{-1}Z'R = HR, \qquad H = Z(Z'Z)^{-1}Z'.

Tukey named H the hat matrix, since it puts the hat on R; Hoaglin and Welsch (1978) made it a regression diagnostic. Row k of H states that each fitted revenue is a weighted average of all observed revenues:

\quad
(3)\qquad
\hat R(k) = \sum_j h_{kj}\,R(j).

For the bivariate regression the weights have a closed form. With deviations x(k)=X(k)-\bar X,

\quad
(4)\qquad
h_{kj} = \frac{1}{N} + \frac{x(k)\,x(j)}{\sum x^2}.

The diagonal element h_{kk}, written h(i,t) for company-year (i,t), is the leverage: the weight a company-year places on its own fitted value. Three properties follow from the algebra of H.

  1. H is symmetric and idempotent (HH=H), so h_{kk}=\sum_j h_{kj}^2 \ge h_{kk}^2, which gives 0\le h_{kk}\le 1. With an intercept, h_{kk}\ge 1/N.
  2. The trace of H equals the number of parameters: \sum h(i,t)=p=2. The average leverage is 2/N.
  3. Leverage depends on X alone. Revenue (dependent variable) does not enter H.

3. Leverage in standard deviations

The diagonal of H has a transparent form once operating cost is standardized. Compute

\begin{aligned}
&(5)\quad &\bar X &= \tfrac{1}{N}\textstyle\sum_{i,t} X(i,t) &&\text{mean}
\\[8pt]
&\quad &S &= \sqrt{\tfrac{1}{N}\textstyle\sum_{i,t}\bigl(X(i,t)-\bar X\bigr)^2} &&\text{standard deviation, divisor } N
\\[8pt]
&\quad &z(i,t) &= \frac{X(i,t)-\bar X}{S} &&\text{standardized deviation}
\end{aligned}

Since \sum x^2 = NS^2,

\quad
(6)\qquad
h(i,t) = \frac{1}{N} + \frac{x(i,t)^2}{NS^2} = \frac{1+z(i,t)^2}{N}.

A company-year at the mean has the minimum leverage 1/N. One standard deviation away doubles it; two standard deviations multiply it by five. The divisor matters: with S computed on N-1 degrees of freedom (Excel STDEV, pandas .std()), the identity becomes h=1/N+z^2/(N-1).

4. Why a high-leverage residual deceives

The residuals are e=(I-H)R=(I-H)u, where u is the vector of true errors. With homoskedastic errors of variance \sigma^2,

\quad
(7)\qquad
\operatorname{Var} e(i,t) = \sigma^2\bigl(1-h(i,t)\bigr).

The residual of a distant company-year has small variance because the line has been pulled toward it. At h=0.5 the residual carries half the error variance; the observation has absorbed the other half into its own fit. The remedy is the deleted residual,

\quad
(8)\qquad
\tilde e(i,t) = \frac{e(i,t)}{1-h(i,t)} = R(i,t) - \hat R_{(-i,t)}(i,t),

the prediction error for (i,t) from the regression estimated without it. The identity is exact and requires no re-estimation.

5. HC3 is a deletion statistic

Deleting company-year k changes the coefficient vector by the Sherman–Morrison identity:

\quad
(9)\qquad
\hat\beta - \hat\beta_{(-k)} = \frac{(Z'Z)^{-1} z_k\, e_k}{1-h_k}.

For the slope, the relevant element of (Z'Z)^{-1}z_k is the weight w_k = x_k/\sum x^2, so

\quad
(10)\qquad
d(i,t) \equiv b - b_{(-i,t)} = \frac{w(i,t)\,e(i,t)}{1-h(i,t)} = w(i,t)\,\tilde e(i,t).

The HC3 covariance estimator of MacKinnon and White (1985), discussed in Davidson and MacKinnon (2004), weights each squared residual by (1-h)^{-2}. For the slope,

\quad
(11)\qquad
V_3 = \sum_{i,t} \frac{w(i,t)^2\,e(i,t)^2}{\bigl(1-h(i,t)\bigr)^2} = \sum_{i,t} d(i,t)^2, \qquad \mathrm{SE}_3 = \sqrt{V_3}.

The HC3 variance of the slope is the sum of the squared slope changes from deleting each company-year in turn. This identity converts a covariance formula into a diagnostic. The share

\quad
(12)\qquad
c(i,t) = \frac{d(i,t)^2}{V_3}, \qquad \sum c(i,t) = 1,

measures how much of the estimated uncertainty one company-year generates. The average share is 1/N. The factor (1-h)^{-2} is 1.44 at h=0.167 and 4.00 at h=0.5: HC3 enlarges the contribution of distant company-years, which is the correction section 4 requires.

6. From company-year to company

A company is not three independent observations. Its years share management, contracts, accounting policy, and cost structure. The question that matters for the arm’s-length range is whether the company determines the slope. Deleting the block of years \mathcal I_i belonging to company i gives the generalization

\quad
(13)\qquad
\hat\beta - \hat\beta_{(-i)} = (Z'Z)^{-1} Z_i' \bigl(I - H_{ii}\bigr)^{-1} e_i,

where Z_i, e_i and H_{ii} are the rows and the diagonal block of H for company i. Its slope element is \Delta(i) = b - b_{(-i)}. In practice, re-estimate the regression without company i; the block formula serves as a check. The company test compares \Delta(i) with the precision of the full-sample slope: a company is intolerable when |\Delta(i)| > \mathrm{SE}_3.

The company test is not redundant with the year-level shares. Year-level effects d(i,t) of opposite sign largely cancel when the company is deleted as a block. The cancellation is not exact arithmetic: \Delta(i) differs from \sum_t d(i,t) because the off-diagonal elements of H_{ii} enter through (I-H_{ii})^{-1}. The symbol Δ(i) means the change in the slope when one company is deleted. It is not a first difference over time like ΔINVT, and it is not a derivative.

7. Boundaries

GradeRule (p = 2)Source
Intolerableh>0.5Huber (1981)
or c>0.5EdgarStat
Influentialc>4/N, four times the average shareBelsley, Kuh and Welsch (1980)
High leverageh>4/N, i.e. |z|>\sqrt3=1.73Hoaglin and Welsch (1978)
Tolerableotherwise
Company intolerable|\Delta(i)|>\mathrm{SE}_3EdgarStat

References

Belsley, D. A., E. Kuh, and R. E. Welsch (1980), Regression Diagnostics: Identifying Influential Data and Sources of Collinearity, Wiley.

Davidson, R., and J. G. MacKinnon (2004), Econometric Theory and Methods, Oxford University Press.

Hoaglin, D. C., and R. E. Welsch (1978), “The Hat Matrix in Regression and ANOVA,” The American Statistician 32(1): 17–22.

Huber, P. J. (1981), Robust Statistics, Wiley.

MacKinnon, J. G., and H. White (1985), “Some Heteroskedasticity-Consistent Covariance Matrix Estimators with Improved Finite Sample Properties,” Journal of Econometrics 29(3): 305–325.