Basic Multivariate Process Monitoring (Section 4.3 of Vardeman and Jobe) 1 Realm of Application • Where relationships between variables matter • Where there will be “automatic” calculation 2 “Obvious” Points About Such Situations • Separate monitoring of variables (e.g. using multiple x charts) can never hope to detect “unusual relationships” … a single summary of all variables is needed • Practical use of an “out of control” signal will require digging into the data to track down what has changed • One must have (and use) some measure of “relationship” between variables 3 Correlation (A Measure of Linear Association Between x and y) • Sample version is rxy = ∑ ( x − x )( y − y ) 2 2 ∑(x − x) ∑( y − y) −1 ≤ rxy ≤ 1 and • We’ll call the “population”/theoretical version ρ xy and − 1 ≤ ρ xy ≤ 1 4 r ≈ −.1 r ≈ .8 r ≈ −.8 5 Use of Correlation (“Standard Relationship”) Information for p Variables?? • Collect correlations and variances in a matrix σ 12 ρ 21σ 1σ 2 Vp× p = M ρ p1σ 1σ p ρ12σ 1σ 2 σ 22 M ρ p 2σ 2σ p L ρ1 pσ 1σ p L ρ 2 pσ 2σ p O M L σ p2 • Plot the summary statistic (4.24) X 2 ' −1 = n x − µ V x − µ p×1 p×1 p× p p×1 p×1 6 More Use of Correlation Information • Use an upper control limit (only) CLX 2 = p UCLX 2 = p + 3 2 p • Large values of the plotted statistic signal – a change in one of the p means or standard deviations, OR – a change in the relationship between at least 2 of the p variables 7 ??? Interpretation ??? • p=2 version of this can be written without matrix notation 2 2 x1 − µ1 x1 − µ1 x2 − µ2 x2 − µ2 1 2 X = − 2 ρ12 + 2 (1 − ρ12 ) σ 1 / n σ1 / n σ 2 / n σ 2 / n • in-control values of this statistic come from vectors of sample means inside appropriate p-dimensional “footballs” centered at the vector of standard means with size and 8 orientation controlled by V Example 4.5 0 n = 1 p = 2 µ = σ 1 = 15 σ 2 = 18 2×1 0 2 2 x x x x X2 = 1 − 1 2 + 2 92.16 144 207.36 ρ12 = .6 “in-control” diameters 9 Retrospective Analyses • Requires the same kind of pooling together of information from r samples as for retrospective x charts • For a single sample (l) s12 ¶ = r21s1s2 Vp× p M rp1s1s p and r12 s1s2 s22 M rp 2 s2 s p x1 x 2 = x M p×1 x p L r1 p s1s p L r2 p s2 s p O M 2 s p L 10 Retrospective Analyses (cont’d) • Sensible pooled estimators of µ and V are p× p p ×1 then r r ¶ µ = ∑ nl x l / ∑ nl p×1 i =1 p ×1 i =1 and r r ¶ µ V = ∑ (nl − 1)V / ∑ (nl − 1) l p× p i =1 p × p i =1 (this is pooling of sample-by-sample information … not pooling of r samples) 11 Workshop Exercise • Consider two (A diameter, B diameter) pairs for Example 4.5 (see slide 9): (25,30) and (25,-30) – plot them on slide 9 – compute X 2 for both of them – determine whether either unit would cause an out-of-control signal for either diameter alone on an individuals chart (an n=1 x chart) 12