When multiple quality characteristics move together, monitoring them one at a time can hide unusual joint behavior and multiply the chance of false alarms across several charts. Multivariate SPC produces one signal that accounts for relationships among correlated critical-to-quality characteristics (CTQs). Hotelling's T-squared is a widely used statistic for this purpose, as explained in the NIST multivariate control chart guidance.
This article explains when multivariate SPC is appropriate, why separate charts can misread correlated CTQs, and how to interpret and investigate a T-squared signal. It also provides an illustrative example, key assumptions, and four Air Academy Associates resources for building practical skills.
Key Takeaways
- Separate charts can miss unusual joint patterns among correlated CTQs.
- Hotelling's T-squared summarizes multivariate distance in one statistic.
- Phase I and Phase II require different control-limit formulas.
- Investigate contributions and process evidence before assigning a cause.
- Remove baseline points only after confirming an assignable cause.
When Multivariate SPC Beats Separate Control Charts for Correlated CTQs

The first decision is whether to use a multivariate control chart or individual univariate charts. Separate charts can be suitable when the characteristics are reasonably independent and each marginal shift matters on its own. When CTQs are correlated—such as tensile strength and elongation in a polymer—individual charts may miss changes in their joint relationship.
Each chart also carries its own false-alarm probability. If two chart signals were independent and each used a 0.27% false-alarm probability, the probability of at least one false alarm would be 1 − (1 − 0.0027)², or about 0.54% per sampling occasion. When the signals are correlated, the actual combined probability depends on that dependence, so simply adding individual rates is only an approximation.
A properly calibrated multivariate chart monitors the CTQs together at one chosen significance level while accounting for their covariance structure. That makes it useful when the process question concerns joint behavior, although univariate charts may still be retained for diagnosis.
| Situation | Recommended Approach | Reason |
|---|---|---|
| CTQs are reasonably independent | Separate univariate charts | Simpler monitoring and direct interpretation |
| CTQs are correlated | Multivariate SPC (T-squared chart) | Uses the covariance structure to assess joint behavior |
| Alarm occurs on T-squared chart | Contribution diagnostics plus process review | Helps identify variables to investigate |
| Assignable cause confirmed in Phase I | Exclude affected baseline point and recompute | Prevents a known special cause from defining the baseline |
How Correlated CTQs Create Misleading Univariate Signals
Picture a machined part where diameter and wall thickness are measured together. If process physics make these dimensions move in a predictable relationship, their joint behavior forms an elliptical region in two-dimensional space rather than two unrelated intervals.
A point can fall within both individual control limits yet lie outside that expected joint region. Conversely, values near the edges of both individual charts can still represent normal joint variation.
This is why a multivariate chart can complement individual charts: it evaluates the relationship among variables, while the univariate views help diagnose which measurements deserve attention. Air Academy's overview of DMAIC data and control tools provides useful context for placing SPC within a broader improvement project.
Hotelling's T-Squared: How the Multivariate Control Chart Statistic Works

Hotelling's T-squared is a covariance-adjusted squared distance from an observation or subgroup mean to the reference mean. The NIST Hotelling chart explanation describes the quadratic form and the multivariate-normal framework behind it.
For individual observations, the core quadratic form is T² = (x − x̄)ᵀ S⁻¹ (x − x̄), where x is the observation vector, x̄ is the reference mean vector, and S⁻¹ is the inverse sample covariance matrix. A larger value indicates greater multivariate distance from the reference center, but the exact scaling and control limit depend on the sampling structure and monitoring phase.
The Hotelling chart implementation notes show how one current R package separates Phase I calibration from Phase II monitoring and reports variable contributions. Contribution measures point to variables worth investigating; they do not, by themselves, prove a root cause.
Illustrative Example: Monitoring Two Correlated CTQs With a Multivariate Control Chart
Note: The values below are illustrative and do not represent a production dataset.
Suppose a plastics extruder tracks melt temperature (X1) and melt pressure (X2). Historical observations yield a reference mean of [215°C, 1,850 psi] and a covariance matrix with a moderate positive correlation (r ≈ 0.65). The appropriate T-squared limit must be selected for individual or subgrouped data and for Phase I or Phase II; finite-sample limits commonly use beta- or F-based formulas rather than one universal cutoff.
At the next sampling point, X1 = 218°C and X2 = 1,790 psi. Assume both values remain within their individual Shewhart limits, while the calculated T-squared value is 9.4 and the applicable UCL is 8.8.
The combination would then signal because elevated temperature paired with lower-than-expected pressure conflicts with the baseline covariance pattern. The example illustrates how separate charts can miss an unusual joint pattern; its numerical UCL is an assumed teaching value, not a reusable limit.
Assumptions and Interpretation Rules for Multivariate SPC Process Monitoring

Hotelling's T-squared charts rely on assumptions about the baseline and sampling process. A violation does not automatically make the chart useless, but it can change the false-alarm rate or weaken interpretation.
1. Multivariate Normality of the CTQs
Classical T-squared limits are derived under an approximately multivariate-normal model. Heavy tails, strong skewness, or influential outliers can make those limits poorly calibrated.
2. Stable Covariance Structure in the Baseline
The Phase I covariance estimate should represent an in-control process. If the covariance structure changes, the reference model no longer describes current joint behavior reliably.
3. Adequate Phase I Data for the Number of Variables
The covariance matrix must be estimable and stable enough for inversion. NIST states the basic requirement p < n − 1 for individual multivariate observations, but practical adequacy depends on the number of variables, subgroup design, data quality, and covariance stability—not a universal 20-to-30-subgroup rule.
4. Independence of Observations Across Subgroups
Classical T-squared charts assume observations or subgroup statistics are independent over time. Autocorrelation can change signal frequency and should be modeled or otherwise addressed before interpreting chart alarms.
5. Correct Control-Limit Formula for the Monitoring Phase
Phase I evaluates historical data used to establish the baseline, while Phase II monitors future observations against that baseline. NIST's phase-specific Hotelling chart formulas distinguish individual observations from subgroups and Phase I from Phase II.
Escalation Rule: What to Do After a T-Squared Alarm in Multivariate SPC
A T-squared alarm indicates unusual joint behavior; it does not identify the cause. First verify data entry, sampling consistency, and the measurement system before assuming the process changed.
Next, review contribution diagnostics and the individual charts to identify variables and time points for investigation. Use process history, materials, setup, tooling, environment, and operator information to test possible causes; correlation alone does not establish causation.
For Phase I baseline refinement, exclude an observation only when an assignable cause is identified and corrected, then recompute and retest the limits. NIST Phase I guidance explicitly ties deletion to an identified special cause, so a signal by itself is not enough.
A trustworthy escalation process also depends on reliable measurements. Air Academy's guide to measurement variation and MSA explains why measurement error should be assessed before statistical conclusions drive corrective action.
Air Academy Associates' Statistical Process Control Short Course covers control-chart selection and interpretation, including how to distinguish common-cause from special-cause variation. Its KISS (Keep It Simple Statistically) approach is designed for practical shop-floor and laboratory use.
Build the Skills to Apply Multivariate SPC: Relevant Courses and Services

Applying multivariate SPC confidently requires practical command of control charts, measurement quality, covariance, and response planning. The following four Air Academy Associates resources build complementary skills for that work.
Statistical Process Control Short Course
This short course builds a foundation in control-chart selection, construction, and interpretation. Participants learn to select charts for common data types, set control limits, and distinguish common-cause from special-cause variation. Explore the SPC Short Course.
Advanced Model Building Short Course
This course develops skills in building and evaluating models for multiple related variables, including regression and response-surface methods. That background helps practitioners examine covariance, collinearity, model assumptions, and potential process drivers after an alarm. Learn about Advanced Model Building.
- Covers regression and multivariable model structures
- Addresses variable selection and collinearity
- Supports interpretation of relationships among process variables
Graphical and Measurement Tools Short Course
This course covers graphical analysis, correlation, and measurement-system evaluation—useful preparation before estimating a covariance matrix or setting monitoring limits. Measurement error can distort variance and covariance estimates, so data quality should be checked before chart interpretation. View Graphical and Measurement Tools.
Lean Six Sigma Black Belt Online Course
Multivariate SPC can support the Control phase of a DMAIC project when several related CTQs must be monitored after improvement. The Black Belt online course places monitoring within a broader project framework that includes measurement-system analysis and control planning. Explore the Black Belt Course.
Conclusion
Multivariate SPC helps monitor joint process behavior when correlated CTQs cannot be interpreted adequately in isolation. Hotelling's T-squared provides one covariance-adjusted signal, while individual charts, contribution diagnostics, and process evidence support diagnosis.
When an alarm occurs, verify the data, investigate likely contributors, confirm any assignable cause, and revise a Phase I baseline only when the evidence supports doing so. A clear control plan and response process helps operators act consistently after monitoring signals.
Air Academy Associates offers Statistical Process Control training that helps teams select, interpret, and respond to control charts in real process settings. Get started with us today.
FAQs
When Should Correlated Quality Characteristics Be Monitored Together?
Monitor characteristics together when their relationship is meaningful to process performance, when a joint shift matters, or when customer requirements depend on combined performance. A multivariate chart can complement individual charts by detecting unusual combinations that marginal limits may miss.
How Is a Multivariate Control-Chart Signal Investigated?
First verify the data, measurement system, and sampling conditions. Then review contribution diagnostics, individual measurements, and recent process changes before using root-cause analysis or designed experiments to test explanations.
Can Separate Stable Charts Still Miss an Unusual Joint Pattern?
Yes. Individual charts can remain within their limits while the correlation or combined pattern among variables changes. Multivariate SPC is designed to detect that type of joint departure.
