Description
Monitor many correlated metrics with one statistic for unusual levels (T-squared) and one for unusual patterns (Q), then use component scores to diagnose the cause.
Separate control charts on correlated metrics inflate false alarms (nine independent charts at 95 percent signal about a third of the time) and miss joint departures. Jackson sets four goals: one answer to whether the process is in control, a specified overall Type I error, account of the correlations, and a route to diagnosis. Hotelling's T-squared, the sum of squared standardised principal component scores, gives the single answer with an F-based limit; a point can sit inside every univariate band yet outside the T-squared ellipse. When fewer components than variables are kept, the residual sum of squares Q tests whether an observation fits the model at all; a large Q flags bad data, instrumentation faults or a changed correlation structure. Only after a signal are component scores or residuals examined to find which metric moved. Subgroup versions split T-squared into mean and dispersion parts.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Collect a stable base period, then compute means and the covariance or correlation matrix of the metrics.
- Run PCA and choose the number of components k with a stopping rule.
- For each new observation compute component scores and T-squared and compare T-squared with its F-based limit.
- Compute Q from the residuals of the k-component reconstruction and compare it with the Jackson-Mudholkar limit.
- Examine T-squared and Q first; only when one signals, inspect component scores or residual contributions to diagnose which metrics moved.
- Watch component charts for trends even when in control, and use Bonferroni limits if plotting several component charts.
- Re-estimate the base period when the correlation structure has genuinely changed, since drift inflates Q.
Sources
- A User's Guide to Principal Components (Jackson, 1991)
- Multivariate Statistical Inference and Applications (Rencher, 1998)
Also known as
- PCA control charts
- MSPC
- Hotelling T-squared chart
- Q statistic
- Squared prediction error (SPE)
- Hawkins statistic
- Control ellipse
Categories
- Analytics & Behaviour
- Statistics & Modelling