Multi-response Control Chart — Multivariate Process Monitoring
Multi-response Statistical Process Control Chart · Also known as: multivariate control chart, multi-response SPC, MRCC, multiple-response monitoring chart
A multi-response control chart simultaneously monitors two or more correlated quality characteristics on a single chart, preserving the correlation structure that univariate charts ignore. Built on Hotelling's T² statistic and its time-weighted extensions (MEWMA, MCUSUM), it detects process shifts that would be missed if each response were charted independently. It is the standard tool in manufacturing and service quality when product performance depends on multiple interrelated outputs.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use a multi-response control chart when two or more quality characteristics are jointly critical and are correlated; monitoring them separately with univariate charts inflates the overall false-alarm rate and can miss joint shifts. It is appropriate in Phase II process monitoring once stable Phase I data exist to estimate the mean vector and covariance matrix. Prefer MEWMA or MCUSUM when small, persistent shifts are the main concern; T² is better for detecting large, sudden shifts. Do NOT use this method when responses are genuinely independent (separate univariate charts are simpler and equally valid), when Phase I sample sizes are very small (fewer than 20 subgroups), or when the number of responses p approaches or exceeds subgroup size n (the covariance matrix becomes singular and T² is undefined).
Strengths & limitations
- Maintains the correct overall false-alarm rate (Type I error) regardless of how many responses are monitored, unlike multiple simultaneous univariate charts.
- Detects process shifts in directions that are invisible to individual Shewhart charts by exploiting the correlation structure among responses.
- MEWMA and MCUSUM extensions provide fast detection of small, sustained mean shifts with controllable ARL properties.
- A single chart communicates multivariate process health concisely, simplifying operator decision-making on the shop floor.
- Decomposition methods (e.g., MTY) allow post-signal diagnosis to identify the offending responses, preserving actionability.
- Requires reliable Phase I estimates of the mean vector and covariance matrix; with fewer than roughly 20–30 rational subgroups, control limits are inaccurate and the chart is unreliable.
- Sensitive to departures from multivariate normality; non-normal data require bootstrap or non-parametric control limits.
- When p is large relative to subgroup size n, the covariance matrix is singular and standard T² cannot be computed without dimension reduction.
- A signal on the chart indicates that something has changed but does not immediately reveal which response or factor is responsible — decomposition adds analytical overhead.
Frequently asked
Why not just run several Shewhart charts — one per response?
Running k independent Shewhart charts each at significance level α inflates the overall false-alarm probability to approximately 1 − (1 − α)^k. With five responses at α = 0.0027 each, the joint false-alarm rate exceeds 1%. Beyond the inflated Type I error, separate charts ignore correlations and can miss shifts that only appear when responses are considered jointly — a shift along a correlated direction may not breach any individual chart's limits while clearly exceeding the multivariate UCL.
When should I choose MEWMA over T²?
T² is most powerful for detecting large, abrupt shifts. MEWMA accumulates information across time through an exponential weight λ (typically 0.05–0.2) and detects small persistent mean shifts much faster, with lower out-of-control ARL. Choose MEWMA when process shifts of half a standard deviation or less are practically important and must be caught quickly. For sudden large shifts or when simplicity is paramount, T² is adequate.
How do I identify which response caused a T² signal?
The Mason–Tracy–Young (MTY) decomposition partitions T² into p independent terms, each attributable to one response conditional on the others. A large conditional T² for a specific response identifies it as the primary contributor to the signal. Software packages such as Minitab and JMP implement this decomposition automatically after a signal is detected.
What Phase I sample size is sufficient?
The general recommendation is at least 20–30 rational subgroups in Phase I, with more needed as p increases. Sullivan and Woodall (1996) showed that small Phase I samples inflate the true false-alarm rate substantially. For p = 5 responses and subgroup size n = 5, at least 50 subgroups may be needed to achieve near-nominal ARL₀ performance. When only limited Phase I data are available, use exact F-distribution limits or bootstrap-calibrated limits rather than the asymptotic chi-squared UCL.
Can this method handle non-normal data?
Standard T² and MEWMA assume multivariate normality. Moderate skewness has limited impact on ARL properties, but heavy-tailed or highly skewed distributions require alternatives: bootstrap control limits, distribution-free multivariate charts (e.g., based on data depth or rank statistics), or transformation to approximate normality before applying classical limits.
Sources
- Hotelling, H. (1947). Multivariate quality control illustrated by the air testing of sample bombsights. In C. Eisenhart, M. W. Hastay, & W. A. Wallis (Eds.), Techniques of Statistical Analysis (pp. 111–184). McGraw-Hill. link ↗
- Lowry, C. A., Woodall, W. H., Champ, C. W., & Rigdon, S. E. (1992). A multivariate exponentially weighted moving average control chart. Technometrics, 34(1), 46–53. DOI: 10.1080/00401706.1992.10485232 ↗
How to cite this page
ScholarGate. (2026, June 3). Multi-response Statistical Process Control Chart. ScholarGate. https://scholargate.app/en/experimental-design/multi-response-control-chart
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Control chartExperimental design↔ compare
- Failure Mode and Effects AnalysisExperimental design↔ compare
- Multi-response Design of ExperimentsExperimental design↔ compare
- Multi-response Response Surface MethodologyExperimental design↔ compare
- Process Capability AnalysisStatistics↔ compare
- Statistical Process ControlExperimental design↔ compare