Describe the construction of a CUSUM chart with Decision interval and Reference value.
10 marks · directive: Describe
The core of a good answer
A tabular CUSUM accumulates deviations of each observation (or subgroup mean) beyond a reference value K = kσ from the target μ₀: C⁺ᵢ = max[0, xᵢ − (μ₀ + K) + C⁺ᵢ₋₁] and C⁻ᵢ = max[0, (μ₀ − K) − xᵢ + C⁻ᵢ₋₁], starting at 0; the process is signalled out of control when either sum crosses the decision interval H = hσ, with k = 0·5 (half the shift to be detected) and h = 4 or 5 the usual design.
An outline written by upsc.help, not an official UPSC answer; every fact checked by a separate review. How we work
What the question asks
The question wants the construction steps of the decision-interval (tabular) CUSUM, the meaning and choice of the two parameters K and H, how a signal is read and how the time of the shift is estimated. Mention the V-mask as the equivalent graphical form and why CUSUM beats a Shewhart chart for small shifts.
Answer outline
Introduction
The cumulative sum chart, introduced by E. S. Page (1954), plots Sₘ = Σᵢ₌₁ᵐ (x̄ᵢ − μ₀); because it uses all past data it detects small sustained shifts (about 0·5σ to 1·5σ) much faster than a Shewhart chart.
Step 1: Set the target and scale
- μ₀ = target or in-control mean; σ = standard deviation of the plotted statistic (σ/√n for subgroup means).
- Decide the shift to detect quickly: μ₁ = μ₀ ± δσ.
Step 2: Reference value K
- K = kσ, usually K = δσ/2, i.e. halfway between μ₀ and μ₁ (k = 0·5 for a 1σ shift).
- Deviations smaller than K are treated as noise and do not build up the sum, so K acts as an allowance or slack.
Step 3: Accumulate one-sided sums
- Upper: C⁺ᵢ = max[0, xᵢ − (μ₀ + K) + C⁺ᵢ₋₁]
- Lower: C⁻ᵢ = max[0, (μ₀ − K) − xᵢ + C⁻ᵢ₋₁]
- C⁺₀ = C⁻₀ = 0. Also record N⁺ and N⁻, the number of consecutive periods the sum has been positive.
Step 4: Decision interval H
- H = hσ, commonly h = 4 or 5. Signal when C⁺ᵢ > H or C⁻ᵢ > H; plot both sums against i with a horizontal line at H.
- With k = 0·5 and h = 5 the two-sided in-control average run length is about 465 and a 1σ shift is detected in about 10 samples on average; h = 4 gives an in-control ARL of about 168 (Markov chain approximation computed in Python).
Step 5: After a signal
- Estimated new mean: μ̂ = μ₀ + K + C⁺ᵢ/N⁺ (upward shift) or μ₀ − K − C⁻ᵢ/N⁻ (downward).
- The shift began about N⁺ (or N⁻) periods before the signal; look for an assignable cause there, correct it and reset the sums.
Step 6: V-mask equivalent
- Plot Sₘ and place a V-shaped mask at the latest point with lead distance d and half-angle θ; a point outside the arms signals. A V-mask with d = H/K and tan θ = K (in plotted units) is equivalent to the tabular CUSUM.
Conclusion
A CUSUM is defined by its reference value K (what size of drift is ignored) and its decision interval H (how much evidence is needed to signal); together they fix the in-control and out-of-control average run lengths and make the chart efficient for small persistent shifts.
Examples and data to cite
- Chemical and pharmaceutical processes where small drifts in a mean concentration matter use CUSUM charts rather than X̄ charts.
- Page (1954, Biometrika) proposed the cumulative sum scheme; Barnard (1959) introduced the V-mask.
Facts to remember
- Tabular CUSUM: C⁺ᵢ = max[0, xᵢ − (μ₀ + K) + C⁺ᵢ₋₁].
- Reference value K = kσ ≈ δσ/2; decision interval H = hσ, h = 4 or 5.
- Signal when a one-sided sum exceeds H.
- CUSUM detects small sustained shifts faster than Shewhart charts, which use only the latest sample.
- Two-sided ARL₀ ≈ 465 for k = 0·5, h = 5.