Rational Subgrouping

Rational Subgrouping (CSSGB BoK VI.A.2 — Understand)

Quick Answer: A rational subgroup is a small set of observations taken under essentially the same conditions so that within-subgroup variation estimates short-term common-cause noise, while between-subgroup differences expose shifts and special causes over time. Good subgroup design is as important as picking the chart type.

Definition

Rational subgrouping is the deliberate grouping of process data so that:

• Units inside a subgroup are as alike as practical with respect to assignable causes (same stream, short time window, same machine/cavity when those are known sources).
• Subgroups are taken across time (or across suspected sources, if studying those separately) so that real process changes show up as differences between subgroup averages (or proportions/counts).

Shewhart’s idea: estimate σ from within the subgroup; build limits that will be crossed when a between-subgroup shift appears.

Purpose of Rational Subgrouping

Without rational subgroups, an X-bar chart may be little better than a random average of unrelated things—and control limits will not mean what textbooks assume.

Within vs Between Variation

Think of total variation as:

Total ≈ within-subgroup (short-term) + between-subgroup (shifts, streams, time effects)

• Within: consecutive parts from the same setup in five minutes → mostly inherent process + measurement noise.
• Between: morning vs afternoon average, Lot A vs Lot B, Operator 1 vs Operator 2 day averages.

X-bar chart watches between (location over time).
R or s chart watches within (dispersion stability).
If you put morning and afternoon parts in the same subgroup of five, the range chart absorbs the shift as “within” noise, limits widen, and the X-bar chart may miss the shift you care about.

Design Principles

1. Homogeneity inside the subgroup

Prefer consecutive units from one machine, one cavity, one line, one batch of material—whatever defines a single process stream for the CTQ.

Bad design: five units randomly picked from five different machines and averaged as one “subgroup.” Special machine differences hide inside R and inflate σ.

Good design: chart each critical machine (or cavity) separately, or stratify knowingly; within each chart, subgroups are consecutive pieces from that stream.

2. Subgroup size (n)

Common industrial choices:

• n = 1 — use I-MR when only one measurement per period is available or units are not naturally batched.
• n = 2 to ~9 — often X-bar and R (R loses efficiency as n grows).
• n ≥ ~10 — often X-bar and s (sample standard deviation uses the data better than R).

Larger n makes X-bar more sensitive to small mean shifts (smaller standard error) but costs more sampling and may stretch the “same conditions” assumption if you need a long time to collect n pieces.

3. Frequency and time order

• Sample often enough to catch shifts before much bad product ships.
• Always preserve time order (or production order).
• Align frequency with process rhythm: after changeovers, each shift, each lot—document the rule.

4. What you want to detect

If the business risk is tool wear over the day, subgroups should be short windows across the day so a trend in X-bar appears.
If the risk is piece-to-piece instability, the range chart and measurement system matter as much as the average.

5. Attribute sampling

For p/np/c/u charts, “subgroup” means the inspection sample or area of opportunity for that period (e.g., 200 units inspected each hour; or all boards in a lot). Keep the sampling plan constant when using np or c; allow varying n only with p or u charts designed for that.

Worked Design Scenarios

Scenario A — Bottling line fill weight
Risk: gradual drift and sudden nozzle clog. Design: every 30 minutes, take n = 5 consecutive bottles from the same filler head (or chart heads separately). Plot X-bar and R. Within-R estimates short-term fill noise; X-bar catches drift or clog-related shift between half-hours.

Scenario B — Five-cavity mold
If one subgroup mixes one piece from each cavity, cavity-to-cavity bias inflates R and blurs cavity-specific special causes. Better: subgroup within cavity, or use multi-vari / separate charts per cavity for critical dimensions.

Scenario C — Low-volume custom welds
Only one weldment per job. Rational subgroup size >1 is artificial. Use I-MR on each successive weldment measurement; moving range estimates short-term variability between successive individuals.

Scenario D — Invoice error rate
Each day inspect a random sample of 100 invoices. That daily sample is the subgroup for a p chart (or np if n stays 100). Do not pool Monday–Friday into one weekly subgroup if you need daily detection—unless the process truly only changes weekly and that is the intended detection window.

Common Subgrouping Mistakes

1. Mixing streams inside n (machines, tools, suppliers)
2. Convenience sampling that is not time-ordered
3. n too large over too long a window so “within” includes real shifts
4. Changing n on charts that assume constant n without switching to p/u
5. Cherry-picking best pieces for the subgroup
6. Recalculating limits every day without a stable baseline policy—limits should represent the intended common-cause system

Relation to Control Chart Construction

• Control limits for X-bar use within-subgroup estimates (R̄/d₂ or S̄/c₄). Rational subgroups make those constants meaningful.
• Capability studies that use within-σ inherit the same subgroup logic (III.F).
• When rational subgroups of size >1 are impossible, individuals charts are the honest alternative—not inventing fake groups of unrelated parts.

Bottom Line for VI.A.2

Rational subgrouping structures data so within-subgroup variation ≈ short-term common cause and between-subgroup differences reveal special causes over time. Design for homogeneity inside the group, sensible n and frequency, clear process streams, and a sampling plan that matches the shifts you must detect. Understand the purpose before you construct X-bar/R/s or attribute charts.