Diagnosing Gaussian Assumptions for Exact Mean Inference

Why Gaussian Diagnostics Matter

Exact finite-sample inference for a mean vector is valuable because its calibration does not rely on a large-sample approximation. That exactness is conditional on a model: the observations must be independent and identically distributed from a multivariate Gaussian population, and the covariance matrix must be positive definite. A plot cannot establish those assumptions, but it can reveal features of the data that make them doubtful.

The relevant question is not whether every variable looks roughly bell-shaped on its own. Multivariate Gaussian shape concerns the joint distribution: its contours are elliptical, and every linear combination of the coordinates is normally distributed. A few unusual observations, curved dependence, clusters, or heavy tails can be hard to spot in separate marginal summaries. Diagnostics therefore combine views of the joint data with distance measures that account for the variables’ scales and correlations.

Plots and Mahalanobis Distances

Let \(X_1,\ldots,X_n\) be independent observations in \(\mathbb{R}^p\), with sample mean \(\bar X\) and sample covariance matrix \(S\), using divisor \(n-1\). Assume \(S\) is positive definite when calculating distances. The squared Mahalanobis distance of observation \(X_i\) from a center \(m\), using a positive-definite covariance matrix \(V\), is

Unlike ordinary Euclidean distance, Mahalanobis distance accounts for scale and correlation. A displacement along a high-variance direction contributes less than the same-sized displacement along a low-variance direction. Correlation also changes which combinations of coordinates count as unusual.

Useful diagnostic views include pairwise scatterplots, marginal normal quantile plots, and plots of projections onto selected directions. Pairwise scatterplots can show curvature, clusters, changing spread, or isolated points. Marginal quantile plots assess each coordinate separately; they can miss a non-Gaussian joint pattern. Projection plots can reveal departures along combinations of variables that are not visible in any one coordinate. No finite collection of plots checks every possible direction, so these views are evidence to interpret, not a certificate of normality.

A distance-based view compares squared distances from a center to reference quantiles of a chi-square distribution with \(p\) degrees of freedom. The reference has a direct justification when the population mean and covariance are known and the observations are Gaussian. When the center and covariance are estimated from the same data, however, the resulting distances do not have that simple reference distribution.

Two Properties That Help Interpret Distance Diagnostics

First, population Mahalanobis distance has a simple Gaussian reference distribution. To see why, let \(\Sigma^{-1/2}\) be the symmetric inverse square root of \(\Sigma\), and define \(Z=\Sigma^{-1/2}(X-\mu)\). A linear transformation of a multivariate Gaussian vector is Gaussian, so \(Z\sim N_p(0,I_p)\). Therefore its coordinates are independent standard normal variables, and

The conclusion follows from the definition of a chi-square random variable as a sum of squares of independent standard normal variables. It gives a principled reference for population distances, while also explaining why estimated distances require care.

Second, distances computed around the sample mean with the sample covariance obey an exact constraint. This is one reason they cannot be treated as independent chi-square observations.

For the proof, write \(R_i=X_i-\bar X\). By the definition \(S=(n-1)^{-1}\sum_i R_iR_i^\mathsf{T}\), the sum of the distances is

The trace identity in the second line follows because each scalar \(R_i^\mathsf{T}S^{-1}R_i\) equals \(\operatorname{tr}(S^{-1}R_iR_i^\mathsf{T})\). The final equality uses that \(I_p\) has \(p\) diagonal entries equal to one. Since all the distances are computed using the same fitted mean and covariance, they are also dependent. A chi-square quantile plot of these internal distances can be a useful shape diagnostic, but its reference line is not an exact independent-sample calibration.

Worked Examples

Influential Observations and Affine Invariance

An observation can be unusual relative to the rest of the data and can also influence the diagnostics used to judge it. A distant point may pull the sample mean toward itself and enlarge the sample covariance in its direction. Those changes can reduce its own internal distance, a phenomenon often called masking. Conversely, a point may make other observations look more unusual by changing the fitted center or shape.

A practical check is to compare the full-data analysis with one in which a flagged observation is omitted. Recompute the mean, covariance, pairwise plots, and distances using the remaining observations. A large change is evidence that the observation is influential for these summaries. It is not a reason to delete the point automatically: first check for recording or measurement errors, then decide whether the observation belongs to the population and model being studied. Leave-one-out distances are diagnostics, not automatically calibrated hypothesis tests.

The geometry of Mahalanobis distance also clarifies why changes of measurement units should not alter the diagnostic. If \(Y_i=AX_i+b\), where \(A\) is an invertible \(p\times p\) matrix and \(b\) is fixed, then \(\bar Y=A\bar X+b\) and \(S_Y=AS_XA^\mathsf{T}\). Since \(S_Y^{-1}=(A^\mathsf{T})^{-1}S_X^{-1}A^{-1}\), substitution gives

Thus internal Mahalanobis distances are invariant under invertible changes of coordinates, including changes of units and rotations. This does not make them immune to influential observations: the equality compares transformed versions of the same data, not different data sets with and without a point.

Implications for Exact Mean Inference

The diagnostics matter because the exact reference distribution for a mean test depends on the joint sampling model, not merely on how well a test statistic can be calculated. A nonsingular sample covariance makes the usual statistic computable; it does not establish Gaussianity. Nor does a tidy pairwise scatterplot prove that all linear combinations are normal. With small samples, departures such as skewness, heavy tails, mixtures, and influential observations can be difficult to diagnose reliably.

Use several views and interpret them together. Look for curved or clustered patterns in pairwise plots, unusual marginal quantile plots, and systematic departures in distance-based views. Investigate observations that strongly affect the mean or covariance. If diagnostics raise concern, state that concern when interpreting exact finite-sample \(p\)-values or confidence regions; consider whether the model is scientifically defensible and whether a suitable alternative analysis is needed. A diagnostic is most useful as a prompt for model scrutiny, not as a mechanical pass/fail gate.

Common Mistake

• Wrong reasoning: “The sample Mahalanobis distances can be compared with independent \(\chi^2_p\) values, and the points inside the 95th percentile prove that the exact mean test is valid.”
• Correction: The chi-square result applies to distances from the true population mean using the true covariance under Gaussian sampling. Internal distances use estimates from the same data, are dependent, and have a fixed sum. Even reassuring diagnostics cannot prove joint normality or guarantee exact calibration; the finite-sample theorem requires its assumptions.