11.1 Statistical Process Control for Variables (X-bar/R and X-bar/s Charts)

Statistical Process Control for Variables ($\bar{X}$-$R$ and $\bar{X}$-$s$ Charts)

Core Principle: Variables control charts monitor continuous quality characteristics (dimensions, temperatures, tensile strengths) by tracking both central tendency (location) and dispersion (spread). Statistical process control ensures that manufacturing and service operations operate in an economically stable state governed solely by chance causes of variation.

Statistical Process Control (SPC) is one of the foundational quantitative methodologies in industrial and systems engineering. Developed by Dr. Walter A. Shewhart, whose 1924 memo at Western Electric (soon part of Bell Telephone Laboratories) introduced the control chart, SPC provides an operational protocol for differentiating between routine background noise and actionable system disturbances. For the NCEES PE Industrial and Systems Examination, engineers must master the mathematical formulation, factor selection, diagnostic interpretation, and underlying statistical assumptions of variables control charts.

---

1. Statistical Foundations of SPC

Every manufacturing and operational process exhibits variation. Shewhart categorized this variation into two fundamentally distinct classes:

1. Common Cause (Chance) Variation: The inherent, natural background noise of a process that remains after all identifiable disturbances have been eliminated. Common cause variation arises from dozens of minor, uncontrollable sources—such as ambient humidity fluctuations, minor electrical line noise, or micro-structural material heterogeneity. A process subject solely to common causes is defined as operating in statistical control; its performance is predictable over time.
2. Special Cause (Assignable) Variation: Intermittent, external, or structural disruptions that alter the underlying probability distribution of the quality metric. Examples include a fractured cutting tool insert, a contaminated raw material lot, an uncalibrated optical micrometer, or an untrained substitute operator. A process exhibiting special causes is out of statistical control; its output is unstable and unpredictable.

Total Process Variation
├── Common Cause (Chance) Variation
│   ├── Inherent to the process architecture
│   ├── Stable, predictable probability distribution
│   └── Addressed only by system-level engineering redesign
└── Special Cause (Assignable) Variation
    ├── Extrinsic disturbances and operational shifts
    ├── Unstable, unpredictable distribution parameters (mean shift or variance expansion)
    └── Addressed by frontline root-cause elimination

The Central Limit Theorem in Variables SPC

Let a continuous random variable $X$ represent an individual quality measurement with population mean $\mu$ and population standard deviation $\sigma$. If we collect random samples of size $n$ and calculate the sample mean $\bar{X}$:

$\bar{X} = \frac{1}{n} \sum_{j=1}^n X_j$

The Central Limit Theorem (CLT) guarantees that the distribution of sample averages approaches a normal distribution as subgroup size $n$ increases, regardless of whether the parent distribution of $X$ is normal, skewed, uniform, or bimodal:

$\bar{X} \sim N\left(\mu, \sigma_{\bar{X}}^2\right) \quad \text{where} \quad \sigma_{\bar{X}} = \frac{\sigma}{\sqrt{n}}$

In industrial practice, subgroup sizes of $n = 4$ or $n = 5$ provide sufficient normal approximation for unimodal parent distributions. This statistical property makes $\bar{X}$-charts remarkably robust against departures from normality in individual parts.

The Economic Basis of $3\sigma$ Control Limits

Shewhart established control limits positioned at exactly three standard errors ($3\sigma_{\bar{X}}$) on either side of the center line ($CL = \mu$):

$UCL = \mu + 3\sigma_{\bar{X}} = \mu + \frac{3\sigma}{\sqrt{n}}$ $LCL = \mu - 3\sigma_{\bar{X}} = \mu - \frac{3\sigma}{\sqrt{n}}$

For a normal distribution, the probability of an in-control point falling outside $3\sigma$ limits (Type I error rate, $\alpha$, or false alarm rate) is:

$\alpha = 2 \times \Phi(-3) = 2 \times 0.00135 = 0.0027 \quad (0.27\% \text{ or } 1 \text{ in } 370.4 \text{ subgroups})$

Setting narrower limits (e.g., $2\sigma$) increases sensitivity to small shifts (reducing Type II error, $\beta$), but floods the shop floor with false alarms ($\alpha = 4.56\%$). Conversely, setting wider limits (e.g., $4\sigma$) reduces false alarms but severely delays detection of genuine process upsets. Shewhart's $3\sigma$ limits represent an empirical economic optimum balancing the cost of investigating false alarms against the cost of producing nonconforming product.

---

2. Principles of Rational Subgrouping

Control charts function effectively only when data collection adheres to rational subgrouping. Rational subgrouping governs how samples are formed to isolate sources of variation:

Two primary subgrouping strategies exist in operational environments:

For standard Shewhart variables monitoring, snapshot sampling of consecutive units is the mandatory protocol. If an engineer mistakenly pools units from multiple machines, cavities, or work shifts into a single subgroup, the within-subgroup variance inflates. This widens the control limits and blinds the chart to genuine process shifts.

---

3. $\bar{X}$ and $R$ Control Charts

When the subgroup size is small ($n \le 10$, and conventionally $n = 4$ or $5$), the subgroup sample range $R$ serves as a computationally simple, highly effective proxy for process dispersion.

Formulation and Center Lines

Given $m$ preliminary rational subgroups, each of size $n$:

• For subgroup $i$, the sample mean is $\bar{X}_i = \frac{1}{n} \sum_{j=1}^n X_{ij}$.
• The sample range is $R_i = X_{i,\max} - X_{i,\min}$.

The grand mean (center line of the $\bar{X}$-chart) and average range (center line of the $R$-chart) are:

$\bar{\bar{X}} = \frac{1}{m} \sum_{i=1}^m \bar{X}_i \qquad \bar{R} = \frac{1}{m} \sum_{i=1}^m R_i$

Estimating Population Standard Deviation ($\sigma$)

The relative range $W = R / \sigma$ is a random variable whose expected value depends solely on sample size $n$. In statistical quality literature and the NCEES Reference Handbook, this expected value is tabulated as factor $d_2$:

$E(R) = d_2 \sigma \implies \hat{\sigma} = \frac{\bar{R}}{d_2}$

Derivation of Control Chart Limits

Substituting $\hat{\sigma} = \bar{R}/d_2$ into the theoretical Shewhart limits for $\bar{X}$ yields:

$UCL_{\bar{X}} = \bar{\bar{X}} + \frac{3\sigma}{\sqrt{n}} = \bar{\bar{X}} + \frac{3}{d_2 \sqrt{n}} \bar{R} = \bar{\bar{X}} + A_2 \bar{R}$ $LCL_{\bar{X}} = \bar{\bar{X}} - \frac{3}{d_2 \sqrt{n}} \bar{R} = \bar{\bar{X}} - A_2 \bar{R}$

Where the NCEES factor $A_2$ is defined analytically as:

$A_2 = \frac{3}{d_2 \sqrt{n}}$

For the range chart, the standard deviation of the relative range $W$ is tabulated as $d_3$, so $\sigma_R = d_3 \sigma = \frac{d_3}{d_2} \bar{R}$. The $3\sigma$ limits for the $R$-chart are:

$UCL_R = \bar{R} + 3\sigma_R = \bar{R} \left(1 + 3\frac{d_3}{d_2}\right) = D_4 \bar{R}$ $LCL_R = \bar{R} - 3\sigma_R = \bar{R} \max\left(0, 1 - 3\frac{d_3}{d_2}\right) = D_3 \bar{R}$

Where $D_4 = 1 + 3\frac{d_3}{d_2}$ and $D_3 = \max\left(0, 1 - 3\frac{d_3}{d_2}\right)$. For $n \le 6$, $D_3 = 0$, meaning the range chart has no active lower control limit.

Standard NCEES Control Chart Factors Table

The standard control chart factors for subgroup sizes $n = 2$ through $10$ are below. Use the table in the NCEES handbook on exam day; the values are the same standard constants.

---

4. $\bar{X}$ and $s$ Control Charts

While the sample range $R$ is easy to compute manually, it utilizes only the two extreme observations ($X_{\max}$ and $X_{\min}$) in each subgroup. For moderate to large sample sizes ($n > 10$, or modern computer-monitored processes with $n \ge 5$), the sample standard deviation $s$ provides an unbiased and statistically more efficient estimator of dispersion.

Formulation and Factors

For subgroup $i$, the sample standard deviation is:

$s_i = \sqrt{\frac{1}{n-1} \sum_{j=1}^n (X_{ij} - \bar{X}_i)^2}$

The expected value of $s$ from a normal population is slightly less than the true standard deviation $\sigma$ due to Jensen's inequality: $E(s) = c_4 \sigma$. Therefore, the unbiased estimator of population dispersion is:

$\hat{\sigma} = \frac{\bar{s}}{c_4} \qquad \text{where} \quad \bar{s} = \frac{1}{m} \sum_{i=1}^m s_i$

The control limits for the sample standard deviation ($s$-chart) are:

$UCL_s = B_4 \bar{s} \qquad LCL_s = B_3 \bar{s} \qquad CL_s = \bar{s}$

Where the factors $B_3$ and $B_4$ are derived from the standard error of $s$ ($\sigma_s = \sigma \sqrt{1 - c_4^2}$):

$B_4 = 1 + \frac{3}{c_4} \sqrt{1 - c_4^2} \qquad B_3 = \max\left(0, 1 - \frac{3}{c_4} \sqrt{1 - c_4^2}\right)$

The control limits for the mean chart using sample standard deviation are:

$UCL_{\bar{X}} = \bar{\bar{X}} + A_3 \bar{s} \qquad LCL_{\bar{X}} = \bar{\bar{X}} - A_3 \bar{s} \qquad \text{where} \quad A_3 = \frac{3}{c_4 \sqrt{n}}$

When subgroup sizes vary across samples ($n_i \neq n$), the $\bar{X}$-$s$ formulation is readily adjusted by calculating sample-specific limits using $A_{3,i}$, $B_{3,i}$, and $B_{4,i}$, whereas the $R$-chart cannot cleanly handle variable subgroup sizes.

---

5. Individual and Moving Range ($I-MR$) Charts

In many modern production and automated inspection environments, rational subgrouping with $n > 1$ is technically or economically infeasible. Common scenarios requiring a subgroup size of $n = 1$ include:

• Automated 100% In-line Inspection: Measurements occur on every continuous part with long intervals between units.
• Destructive Testing: Testing a specimen destroys high-value components (e.g., tensile testing of titanium alloy forgings or crash-testing chassis).
• Chemical and Process Industries: A batch reaction in a 10,000-liter bioreactor or chemical distillation column yields a single homogeneous concentration measurement per batch; taking multiple samples from the same tank measures only measurement error, not process variation.
• Operational and Financial Metrics: Tracking daily production yield, monthly utility consumption, or weekly inventory turns.

Formulation of the $I-MR$ Chart

For individual measurements $X_1, X_2, \dots, X_m$, dispersion is estimated by computing the moving range ($MR$) between consecutive observations:

$MR_i = |X_i - X_{i-1}| \quad \text{for } i = 2, 3, \dots, m$ $\bar{MR} = \frac{1}{m-1} \sum_{i=2}^m MR_i$

Because each moving range is computed from two consecutive points, the effective sample size is $n = 2$. From the factor table, $d_2 = 1.128$ and $D_4 = 3.267$. The estimated process standard deviation is:

$\hat{\sigma} = \frac{\bar{MR}}{d_2} = \frac{\bar{MR}}{1.128}$

The control limits for the Individuals ($I$ or $X$) chart are:

$UCL_X = \bar{X} + 3 \frac{\bar{MR}}{d_2} = \bar{X} + 2.660 \bar{MR}$ $LCL_X = \bar{X} - 3 \frac{\bar{MR}}{d_2} = \bar{X} - 2.660 \bar{MR}$ $CL_X = \bar{X} = \frac{1}{m} \sum_{i=1}^m X_i$

The control limits for the Moving Range ($MR$) chart are:

$UCL_{MR} = D_4 \bar{MR} = 3.267 \bar{MR} \qquad LCL_{MR} = D_3 \bar{MR} = 0 \qquad CL_{MR} = \bar{MR}$

---

6. Out-of-Control Detection and Run Rules

A process can become unstable without a single point breaching the $3\sigma$ limits. Systematic patterns, trends, and non-random clustering indicate assignable causes operating within the process.

To standardize pattern recognition, Shewhart charts partition the space between the center line and each control limit into three symmetrical one-sigma zones:

+3 Sigma ---------------------------------------------------- UCL
               Zone A (2-sigma to 3-sigma upper)
+2 Sigma ----------------------------------------------------
               Zone B (1-sigma to 2-sigma upper)
+1 Sigma ----------------------------------------------------
               Zone C (Center line to 1-sigma upper)
Center Line ================================================= CL
               Zone C (Center line to 1-sigma lower)
-1 Sigma ----------------------------------------------------
               Zone B (1-sigma to 2-sigma lower)
-2 Sigma ----------------------------------------------------
               Zone A (2-sigma to 3-sigma lower)
-3 Sigma ---------------------------------------------------- LCL

The Western Electric Company (WECO) Rules

The Western Electric rules flag an assignable cause when any of the following statistical conditions occur:

1. Rule 1 (Extreme Point): 1 point falls beyond Zone A (outside $UCL$ or $LCL$). Probability under pure chance: $p = 0.0027$.
2. Rule 2 (Zone A Alert): 2 out of 3 consecutive points fall in Zone A or beyond on the same side of the center line. Signals a moderate shift in process mean ($\Delta \mu \approx 1.0\sigma$ to $1.5\sigma$).
3. Rule 3 (Zone B Alert): 4 out of 5 consecutive points fall in Zone B or beyond on the same side of the center line. Signals a small shift in process mean ($\Delta \mu \approx 0.5\sigma$ to $1.0\sigma$).
4. Rule 4 (Run Above/Below Center Line): 8 consecutive points fall on the same side of the center line (in Zone C or beyond). Probability under pure chance: $(0.5)^8 = 0.0039$. Signals a persistent process bias.

Nelson Run Rules for Pattern Analysis

Lloyd S. Nelson extended these heuristics to eight formalized detection tests, adding specific pattern signatures:

• Rule 5 (Trend): 6 consecutive points steadily increasing or steadily decreasing. Root cause: progressive tool wear, chemical depletion, or thermal accumulation.
• Rule 6 (Stratification / Hugging Center Line): 15 consecutive points within Zone C (both sides of center line). Root cause: incorrect subgrouping formed by mixing distinct product streams or overestimating process variation.
• Rule 7 (Systematic Oscillation): 14 consecutive points alternating up and down. Root cause: alternating raw material lots, two operators alternating parts, or process over-adjustment (tampering).
• Rule 8 (Mixture / Avoidance of Center Line): 8 consecutive points on both sides of center line with none falling in Zone C. Root cause: blending outputs from two parallel machines with different calibration setpoints.

---

7. Comprehensive Worked Numerical Problem: $\bar{X}$ and $R$ Design

An industrial quality engineering team at an aerospace propulsion facility monitors the outer diameter of a precision fuel injector sleeve using an automated multi-spindle lathe. The engineering design specification is $50.000 \pm 0.300\text{ mm}$.

A Phase I retrospective study collects $m = 25$ rational subgroups, each comprising $n = 5$ consecutive machined sleeves. The preliminary statistical analysis yields the following summary values:

$\sum_{i=1}^{25} \bar{X}_i = 1251.250\text{ mm} \qquad \sum_{i=1}^{25} R_i = 10.500\text{ mm}$

During review of the historical data, Subgroup 14 is flagged with an observed range of $R_{14} = 0.950\text{ mm}$ and an average of $\bar{X}_{14} = 50.120\text{ mm}$. Physical inspection reveals that a cracked carbide cutting insert caused severe chattering during Subgroup 14's cycle. The insert was replaced immediately. The team must eliminate this assignable cause and establish revised operational Phase II control limits.

Step 1: Preliminary Control Limits and Out-of-Control Confirmation

Compute initial grand mean and average range across all 25 subgroups:

$\bar{\bar{X}}_0 = \frac{1251.250}{25} = 50.050\text{ mm} \qquad \bar{R}_0 = \frac{10.500}{25} = 0.420\text{ mm}$

From the NCEES factor table for $n = 5$: $A_2 = 0.577, \quad D_3 = 0, \quad D_4 = 2.114, \quad d_2 = 2.326$

Calculate preliminary limits for the range chart: $UCL_{R,0} = D_4 \bar{R}_0 = 2.114 \times 0.420 = 0.8879\text{ mm}$ $LCL_{R,0} = D_3 \bar{R}_0 = 0\text{ mm}$

Comparing Subgroup 14's range ($R_{14} = 0.950\text{ mm}$) to the upper limit confirms that $R_{14} > UCL_{R,0}$ ($0.950 > 0.8879$). Because an assignable cause was identified (cracked insert), Subgroup 14 must be purged from the baseline.

Step 2: Recalculate Revised Center Lines

Subtract the values of Subgroup 14 from the baseline totals ($m = 24$):

$\sum \bar{X}_{\text{rev}} = 1251.250 - 50.120 = 1201.130\text{ mm}$ $\sum R_{\text{rev}} = 10.500 - 0.950 = 9.550\text{ mm}$

Compute the revised center lines:

$\bar{\bar{X}}_{\text{rev}} = \frac{1201.130}{24} = 50.0471\text{ mm}$ $\bar{R}_{\text{rev}} = \frac{9.550}{24} = 0.3979\text{ mm}$

Step 3: Establish Operational Phase II Control Limits

Calculate the revised range chart limits:

$UCL_R = D_4 \bar{R}_{\text{rev}} = 2.114 \times 0.3979 = 0.8412\text{ mm}$ $LCL_R = D_3 \bar{R}_{\text{rev}} = 0\text{ mm}$ $CL_R = \bar{R}_{\text{rev}} = 0.3979\text{ mm}$

Calculate the revised mean chart limits:

$UCL_{\bar{X}} = \bar{\bar{X}}_{\text{rev}} + A_2 \bar{R}_{\text{rev}} = 50.0471 + (0.577 \times 0.3979) = 50.0471 + 0.2296 = 50.2767\text{ mm}$ $LCL_{\bar{X}} = \bar{\bar{X}}_{\text{rev}} - A_2 \bar{R}_{\text{rev}} = 50.0471 - 0.2296 = 49.8175\text{ mm}$ $CL_{\bar{X}} = \bar{\bar{X}}_{\text{rev}} = 50.0471\text{ mm}$

Step 4: Estimate Process Standard Deviation ($\hat{\sigma}$)

The underlying within-subgroup process standard deviation is:

$\hat{\sigma} = \frac{\bar{R}_{\text{rev}}}{d_2} = \frac{0.3979}{2.326} = 0.1711\text{ mm}$

Engineering Summary: The lathe process exhibits stable dispersion at $\hat{\sigma} = 0.1711\text{ mm}$. The Phase II monitoring protocol uses $UCL_{\bar{X}} = 50.277\text{ mm}$, $LCL_{\bar{X}} = 49.818\text{ mm}$, and $UCL_R = 0.841\text{ mm}$.

---

8. Common PE Exam Pitfalls for Variables Charts

1. Conflating Control Limits with Specification Limits: Control limits reflect the statistical voice of the process ($CL \pm 3\sigma/\sqrt{n}$) calculated exclusively from sample data. Specification limits ($USL, LSL$) reflect the voice of the customer defined by engineering drawings. Never plot specification limits on an $\bar{X}$-chart, and never evaluate process capability directly from control limits.
2. Failing to Verify Dispersion Control First: On exam items presenting both mean and range data, always check whether any point breaches $UCL_R$ or $LCL_R$. If the range chart is out of control, interpreting the $\bar{X}$-chart is a critical methodological error.
3. Confusing Process Spread ($\sigma$) with Sample Mean Spread ($\sigma_{\bar{X}}$): The standard deviation of the population is estimated as $\hat{\sigma} = \bar{R}/d_2$. The standard deviation of the subgroup means is $\sigma_{\bar{X}} = \hat{\sigma}/\sqrt{n} = \bar{R}/(d_2 \sqrt{n})$. Dividing $\bar{R}/d_2$ by an extra $\sqrt{n}$ when asked for process dispersion is an extremely common arithmetic mistake.
4. Misapplying Factors Across Chart Types: Ensure you use $A_2, D_3, D_4$ when using sample ranges ($R$), and switch to $A_3, B_3, B_4$ when given sample standard deviations ($s$). Using $A_2$ with $\bar{s}$ causes immediate failure.
5. Assuming $I$-Charts Are Protected by CLT: Remember that Individual ($I$) charts operate on raw individual measurements ($n=1$), meaning non-normal data will produce substantial false alarms under Shewhart $3\sigma$ formulas.