Robust Distance Cutoffs for Multivariate Diagnostics

Why a Distance Needs a Reference Model

A robust multivariate distance measures how far an observation lies from a fitted center after accounting for the scale and direction of variation. It is useful for identifying observations that are unusual relative to a fitted multivariate model. But a distance by itself does not say how unusual an observation is: that requires a reference distribution and a cutoff.

The familiar chi-square cutoff is justified under a multivariate normal reference model when the location and scatter estimates are suitably consistent. It is not a universal threshold for every robust distance or every data distribution. Heavy tails, mixtures, finite samples, and the details of the fitting procedure can all change the probability of crossing a cutoff.

Definition: Let $X_i\in\mathbb{R}^p$ be an observation, and let $\hat{\mu}_R$ and $\hat{\Sigma}_R$ be robust estimates of location and positive-definite scatter. The squared robust Mahalanobis distance is $$ D_{Ri}^2=(X_i-\hat{\mu}_R)^\mathsf{T}\hat{\Sigma}_R^{-1}(X_i-\hat{\mu}_R). $$ A cutoff $c$ produces a diagnostic flag when $D_{Ri}^2>c$. The flag is relative to the stated estimates and reference calibration; it is not, by itself, a formal hypothesis test.

The scatter matrix must use the same scale convention as the reference distribution. For example, if a robust procedure estimates a scatter matrix that is not calibrated to estimate the population covariance, a normal-theory cutoff cannot automatically be applied as though it were. Any consistency correction required by the estimator is part of the calibration.

Normal-Reference Calibration

If a new observation $X$ follows $N_p(\mu,\Sigma)$, where $\Sigma$ is positive definite, and the true parameters are used, then its squared Mahalanobis distance has a chi-square distribution with $p$ degrees of freedom. To see why, write $Z=\Sigma^{-1/2}(X-\mu)$. Then $Z\sim N_p(0,I)$, and $$ (X-\mu)^\mathsf{T}\Sigma^{-1}(X-\mu)=Z^\mathsf{T}Z=\sum_{j=1}^p Z_j^2\sim\chi_p^2. $$ This fact motivates the cutoff $c=\chi^2_{p,1-\alpha}$, where $\chi^2_{p,1-\alpha}$ denotes the $1-\alpha$ quantile of the chi-square distribution with $p$ degrees of freedom.

Theorem: Suppose $X_1,\ldots,X_n$ are iid $N_p(\mu,\Sigma)$, with fixed $p$ and positive-definite $\Sigma$. Let $\hat{\mu}_R\xrightarrow{p}\mu$ and $\hat{\Sigma}_R\xrightarrow{p}\Sigma$, with $\hat{\Sigma}_R$ positive definite with probability tending to one. For an independent $X_{\mathrm{new}}\sim N_p(\mu,\Sigma)$, the distance computed from these estimates satisfies $D_{R,\mathrm{new}}^2\xrightarrow{d}\chi_p^2$.

Proof. Since $\Sigma$ is positive definite and $\hat{\Sigma}_R\xrightarrow{p}\Sigma$, continuity of matrix inversion at $\Sigma$ gives $\hat{\Sigma}_R^{-1}\xrightarrow{p}\Sigma^{-1}$. Also, $X_{\mathrm{new}}-\hat{\mu}_R$ converges in distribution to $X_{\mathrm{new}}-\mu$: the estimation error $\hat{\mu}_R-\mu$ converges in probability to zero, and $X_{\mathrm{new}}$ has the same distribution for every $n$. Applying Slutsky's theorem and the continuity of the quadratic-form map gives $$ D_{R,\mathrm{new}}^2 \xrightarrow{d} (X_{\mathrm{new}}-\mu)^\mathsf{T}\Sigma^{-1}(X_{\mathrm{new}}-\mu). $$ The limiting quadratic form has the $\chi_p^2$ distribution by the standard normal transformation above. $\square$

This result concerns an independent new observation and an asymptotic calibration. It does not say that distances computed for the same observations used to fit location and scatter have an exact chi-square distribution in a finite sample. Those distances share the fitted estimates, and their joint behavior depends on the estimator and the fitting procedure. If the intended diagnostic concerns in-sample distances or a small sample, calibration should reflect that procedure rather than assume an exact chi-square law.

Worked Example: Computing a Distance and Applying a Normal Cutoff

Suppose a robust fit gives $\hat{\mu}_R=(0,0)^\mathsf{T}$ and $$ \hat{\Sigma}_R= \begin{pmatrix} 4&1\\ 1&2 \end{pmatrix}. $$ For the observation $x=(5,3)^\mathsf{T}$, compute the squared robust distance and compare it with the 5% chi-square cutoff for $p=2$.

The determinant of the scatter matrix is $4(2)-1(1)=7$, so $$ \hat{\Sigma}_R^{-1}=\frac{1}{7} \begin{pmatrix} 2&-1\\ -1&4 \end{pmatrix}. $$ The displacement is $x-\hat{\mu}_R=(5,3)^\mathsf{T}$. Therefore, $$ D_R^2 = \begin{pmatrix}5&3\end{pmatrix} \frac{1}{7} \begin{pmatrix}2&-1\\-1&4\end{pmatrix} \begin{pmatrix}5\\3\end{pmatrix} = \frac{1}{7}\left(50-30+36\right)=8. $$ The 95th percentile of $\chi_2^2$ is approximately $5.991$. Since $8>5.991$, the observation is flagged under this normal-reference rule. The conclusion is conditional on the cutoff being appropriate for the fitted scatter and the intended diagnostic; it does not establish that the observation is erroneous or that a formal null hypothesis has been rejected.

How Nonnormal References Change the Cutoff

Robustness of the location and scatter estimates does not make the distribution of the distance distribution-free. If the data have heavier tails than a normal distribution, observations can exceed a normal-based cutoff more often than the nominal rate. The reference model matters even when the fitted distance uses robust estimates.

Result: Let $X$ have a multivariate $t$ distribution with $\nu>2$ degrees of freedom, location $\mu$, and covariance matrix $\Sigma$. Represent it as $$ X=\mu+\sqrt{\frac{\nu-2}{W}}\,\Sigma^{1/2}Z, $$ where $Z\sim N_p(0,I)$, $W\sim\chi_\nu^2$, and $Z$ and $W$ are independent. Then the squared distance using the true $\mu$ and $\Sigma$ has distribution $$ (X-\mu)^\mathsf{T}\Sigma^{-1}(X-\mu) \sim \frac{p(\nu-2)}{\nu}F_{p,\nu}. $$

Derivation. Substitution into the quadratic form yields $$ (X-\mu)^\mathsf{T}\Sigma^{-1}(X-\mu) =\frac{\nu-2}{W}Z^\mathsf{T}Z. $$ By definition, $Z^\mathsf{T}Z\sim\chi_p^2$, independently of $W$. The ratio $(Z^\mathsf{T}Z/p)/(W/\nu)$ has the $F_{p,\nu}$ distribution. Rearranging gives the stated factor $p(\nu-2)/\nu$. Thus a chi-square cutoff is not an exact $t$-reference cutoff, even though $\Sigma$ is the covariance matrix. $\square$

Worked Example: A Multivariate t Reference

Take $p=2$, $\nu=5$, and use the normal-reference cutoff $c=5.991$. Under the multivariate $t$ model above, the distance has distribution $\frac{2(5-2)}{5}F_{2,5}=1.2F_{2,5}$. Therefore, $$ P(D^2>5.991) = P\left(F_{2,5}>\frac{5.991}{1.2}\right) = P(F_{2,5}>4.9925). $$ For $F_{2,5}$, the upper-tail probability at $x$ is $\left(5/(5+2x)\right)^{5/2}$. Hence $$ P(D^2>5.991) = \left(\frac{5}{5+2(4.9925)}\right)^{5/2} \approx 0.0643. $$ The normal-based 5% cutoff therefore flags about 6.4% under this $t_5$ reference, rather than 5%. If the target flag rate is 5% under the $t_5$ model, its cutoff should instead be the 95th percentile of $1.2F_{2,5}$.

A mixture can create a similar discrepancy, and its tail probability can be calculated directly when the component distributions are specified. The following example illustrates the calculation for a two-dimensional mixture and a particular fitted scatter matrix. It does not claim that every robust estimator will return that matrix for this mixture.

Worked Example: A Gaussian Mixture and Its Flag Rate

Suppose a diagnostic uses $\hat{\mu}_R=0$ and $\hat{\Sigma}_R=I$, and the reference distribution is a centered mixture: with probability $0.9$, $X\sim N_2(0,(1/1.8)I)$; with probability $0.1$, $X\sim N_2(0,5I)$. Find the probability that $D_R^2>5.991$.

Within a component with covariance $sI$, the squared distance $X^\mathsf{T}X$ has distribution $s\chi_2^2$. For a $\chi_2^2$ variable, the upper-tail probability at $y$ is $e^{-y/2}$. Thus, conditional on the first component, the exceedance probability is $$ P(D_R^2>5.991\mid \text{component 1}) = \exp\left(-\frac{1.8(5.991)}{2}\right) = e^{-5.3919}. $$ For the second component it is $$ P(D_R^2>5.991\mid \text{component 2}) = \exp\left(-\frac{5.991}{2(5)}\right) = e^{-5.991/10}. $$ Weighting by the mixture probabilities gives $$ P(D_R^2>5.991) = 0.9e^{-5.3919}+0.1e^{-0.5991} \approx 0.0590. $$ The mixture's exceedance probability is about 5.9%, not 5%. This calculation also shows why the mixture weights and component scales must both appear in the calibration.

Choosing and Reporting a Cutoff

A defensible cutoff begins with a clear statement of the target reference and the kind of distance being assessed. Under a normal model, a chi-square quantile can be a useful asymptotic reference when robust location and scatter estimates are consistent for the normal location and covariance. Under a specified nonnormal model, derive or simulate the corresponding distance distribution instead. If finite-sample in-sample distances are the target, the simulation should reproduce the fitting procedure and calculate the same type of distances.

A parametric calibration can proceed by simulating data sets from the reference model, fitting the robust location and scatter in each simulated data set, and recording the diagnostic distances of interest. A chosen quantile of those simulated distances provides a cutoff for that specified procedure and sample size. If the target is a future observation, simulate an independent future observation in every replicate; if the target is an in-sample distance, reproduce the in-sample calculation. These are different calibration targets.

A diagnostic flag and a formal test answer different questions. A flag marks an observation whose distance exceeds a reference cutoff. It does not, on its own, specify a null hypothesis about the full data-generating process, account for the number of observations inspected, or establish a desired familywise error rate. A formal test requires a stated null model and a calibration of the relevant test statistic under that null. Report the distance, the cutoff, the reference model, and whether the calibration is approximate, simulated, or exact.

Common Pitfall

Wrong reasoning: “The distance uses robust estimates, so the $\chi_p^2$ cutoff is valid even when the data are heavy-tailed.”
Correction: Robust estimation limits sensitivity to some departures in fitting location and scatter; it does not force the observations' radial distances to follow a chi-square distribution. The cutoff still depends on the reference distribution, scale convention, sample size, and whether distances are in-sample or for new observations.
Wrong reasoning: “A point above the cutoff has failed a formal 5% test.”
Correction: It has been flagged by a diagnostic rule calibrated to a stated reference. A formal test and its error rate require a specified null hypothesis and calibration for the test actually being conducted.

Key takeaway: Robust Mahalanobis distance supplies a geometric measure of departure from a fitted center. The cutoff supplies a probabilistic interpretation, and that interpretation is valid only relative to its stated reference model and calibration target.

Check Your Understanding

Use the definitions and calculations above to check your understanding.

What assumptions make the asymptotic chi-square result apply to the distance of an independent new observation?
For a multivariate $t$ reference with covariance $\Sigma$, why is the squared distance not distributed as $\chi_p^2$?
In the mixture example, what role do the component weights play in the exceedance probability?
Why might an in-sample distance need a different calibration from the distance of a new observation?
State one difference between a diagnostic flag and a formal hypothesis test.