Hotelling's T-squared Chart

The Hotelling's T-squared statistic is the core calculation, representing the squared statistical distance (Mahalanobis distance) of a sample's mean vector from the in-control process mean vector. It condenses multiple correlated variables into a single monitoring statistic.

• Formula: (T^2 = n(\bar{x} - \mu_0)' S^{-1} (\bar{x} - \mu_0))
• (\bar{x}): Sample mean vector.
• (\mu_0): Historical in-control process mean vector.
• (S^{-1}): Inverse of the sample covariance matrix.
• A single plotted point on the T-squared chart represents the combined status of all p variables for that sample.

This is the chart's defining advantage over multiple univariate charts. It uses the covariance matrix (S) in its calculation, which captures the linear relationships between all monitored variables.

• Key Benefit: Correctly identifies out-of-control signals that would be missed by separate charts. For example, two variables might each be within their individual control limits, but their specific combination may be improbable given their typical correlation, triggering a T-squared alert.
• Avoids False Alarms: Reduces the overall probability of a Type I error (false out-of-control signal) compared to running many independent univariate charts.

The Upper Control Limit for a Hotelling's T-squared chart is based on the F-distribution, reflecting the multivariate nature of the data.

• Formula for Phase I (Retrospective Analysis): (UCL = \frac{(m-1)(n-1)p}{m(n-1)-p+1} F_{(\alpha, p, m(n-1)-p+1)})
• Formula for Phase II (Process Monitoring): (UCL = \frac{p(m+1)(n-1)}{m(n-1)-p+1} F_{(\alpha, p, m(n-1)-p+1)})
• m: Number of historical subgroups used to estimate (\mu_0) and (S).
• n: Subgroup size.
• p: Number of quality characteristics (variables).
• (F_{(\alpha, p, v)}): The upper (\alpha) percentage point of the F-distribution with p and v degrees of freedom. There is only an UCL; the lower limit is zero.

A point exceeding the UCL signals a multivariate out-of-control condition, but it does not indicate which variable(s) caused it. Diagnosis requires supplemental procedures:

• MYT Decomposition (Mason, Young, Tracy): A method to decompose the overall (T^2) value into independent components attributable to each variable, both unconditionally and conditional on others.
• Bonferroni-adjusted Univariate Charts: Plotting the individual variables on charts with adjusted, tighter control limits (e.g., using (\alpha/p)) to see which ones appear unusual.
• Principal Component Analysis (PCA) Charts: Monitoring the process via the principal components of the covariance matrix can sometimes isolate the source of variation.

The application of Hotelling's T-squared charts is formally split into two distinct phases with different objectives and calculations.

• Phase I (Retrospective): Goal: Establish a stable, in-control baseline. Historical data is analyzed to estimate (\mu_0) and (S), identify and remove special causes, and finalize the control limits for ongoing monitoring.
• Phase II (Monitoring): Goal: Detect shifts from the established baseline. New samples are plotted against the limits derived from the finalized Phase I estimates. The UCL formula changes because the parameters are now treated as known constants from the historical dataset.

Hotelling's T-squared chart is powerful but has specific requirements for valid use.

Typical Applications:

• Monitoring multiple dimensions of a manufactured part (e.g., length, width, thickness).
• Tracking correlated chemical concentrations in a batch process.
• Observing multiple sensor readings from a complex machine.

Critical Prerequisites:

1. Multivariate Normality: The underlying process data for the p variables should follow a multivariate normal distribution.
2. Rational Subgrouping: Samples must be collected in rational subgroups where within-subgroup variation is due to common causes only.
3. Adequate Historical Data (Phase I): A sufficiently large, stable historical dataset (m subgroups) is required to reliably estimate the mean vector and covariance matrix.

Multivariate Statistical Process Control (MSPC) is the extension of univariate SPC to processes with multiple, correlated quality characteristics. It uses multivariate statistics, like the T-squared statistic, to monitor the mean vector and covariance structure of the process. Unlike monitoring variables individually, MSPC accounts for correlations, preventing missed signals and false alarms. It is essential in complex manufacturing and data pipelines where outputs are interdependent.

A control chart is a time-series graph with a central line (process mean) and upper/lower control limits. It is the primary tool in Statistical Process Control for distinguishing common cause from special cause variation. Hotelling's T-squared chart is a multivariate control chart that plots the Mahalanobis distance of a sample's mean vector from the target mean vector. Its control limit is a single upper bound, as the T-squared statistic is always non-negative.

The Mahalanobis distance is a multivariate measure of how many standard deviations a point is from the mean of a distribution, accounting for correlations between variables. It is the core calculation for the T-squared statistic (T² = D²). Unlike Euclidean distance, it scales and rotates the data space based on the covariance matrix, making it unitless and sensitive to the underlying data structure. A large Mahalanobis distance indicates an observation is an outlier relative to the correlated process behavior.

Principal Component Analysis (PCA) is a dimensionality reduction technique often used with Hotelling's T-squared charts. It transforms correlated variables into uncorrelated principal components. Monitoring these components with individual T² charts and a complementary Q statistic (SPE chart) helps diagnose whether a signal is due to a shift in the model space (T²) or a novel type of variation not captured by the model (Q). This PCA-T² approach is standard for high-dimensional process monitoring.

The multivariate normal distribution is the foundational statistical assumption for the standard Hotelling's T-squared chart. It generalizes the normal distribution to multiple dimensions, defined by a mean vector and a covariance matrix. The theoretical control limits for the T-squared chart are derived from this distribution (using the F-distribution). While the chart is robust to minor deviations, severe non-normality can affect the Average Run Length (ARL) and require data transformation or non-parametric alternatives.

Deploying a T-squared chart involves two distinct phases:

• Phase I (Retrospective Analysis): Historical, "in-control" data is used to estimate the process mean vector and covariance matrix and establish the initial control limits. This phase identifies and removes special causes to achieve a stable baseline.
• Phase II (Prospective Monitoring): The chart is used for real-time monitoring of new observations, using the parameters and limits fixed in Phase I. Signals in Phase II indicate the process may have shifted from its established in-control state.