12 SPC Practical Questions Answered — Control Limits, Subgroups, Cpk/Ppk, and Out-of-Control Criteria, All Clarified at Once

Introduction: These 12 Questions Almost Always Appear in Every SPC Training Session

After each SPC training, the questions raised during the Q&A session are highly concentrated. They may seem scattered, but they essentially boil down to these 12 questions, most of which revolve around the same issue: not distinguishing between "how the process is fluctuating" and "what the product requirements are."

This article categorizes these 12 questions into seven themes, providing a straightforward conclusion for each, followed by a detailed explanation. At the end, a quick reference table is included, which can be printed and posted next to the SPC board.

For a comprehensive understanding of the knowledge background and methodology behind these 12 questions, see the SPC Statistical Process Control Comprehensive Guide — a complete knowledge map from control chart selection to process capability, along with accompanying training materials and practical templates.

1. Control Limits and Specification Limits

Q1 | What is the Difference Between Control Limits and Specification Limits? Why Can't We Use Specification Limits as Control Limits?

Conclusion: Control limits are calculated from the process itself, while specification limits are provided by the customer or the drawing. One comes from the process, the other from the requirements.

Control limits are calculated from 25 subgroups of data, answering the question, "How is the process actually fluctuating now?" Therefore, if the data changes, the control limits need to be recalculated. Specification limits are the boundaries that the product must meet, answering the question, "Is this part qualified?" They are set by people and do not change with the process.

A key inference here is: Points on the control chart falling within the limits only indicate that the process is stable, not that the product is qualified.

• When control limits are wider than specification limits, the process is stable, but the product can still exceed the limits.
• When control limits are very narrow, a point may exceed the control limit but still be far from the specification limit, meaning the product is still qualified.

The correct approach is to plot the specification limits as reference lines using dashed red lines, allowing you to easily see the remaining margin between the process and the specifications. Do not use specification limits as control limits to identify anomalies — this can both miss real signals and misidentify normal fluctuations as anomalies.

Use control limits to determine process stability and specification limits to determine product qualification. These two criteria cannot be used interchangeably.

2. Subgroups and Sampling

Q2 | How Should Subgroups Be Divided? Is It Reasonable to Take 5 Pieces Continuously from the Same Shift?

Conclusion: It is reasonable, but only when you are monitoring short-term variations.

The essence of subgroup division is to decide which segment of variation you want to monitor.

Taking five pieces continuously, with almost no time difference and conditions remaining largely unchanged, reflects the instantaneous precision of the equipment, excluding drifts due to shift changes, batch changes, or material changes — this is correct for calculating control limits.

However, if you want to know whether the process is stable over an 8-hour shift, you cannot just take five pieces continuously. You should divide by time, for example, taking five pieces every hour to form a subgroup, allowing the drift between shifts to be revealed.

Three principles:

1. Conditions within the same subgroup must be consistent — same batch, same machine, same method.
2. Subgroups should cover the sources of variation you want to monitor — if you want to manage shift changes, let the subgroups span shifts.
3. Do not mix different conditions within a subgroup — if a subgroup includes both morning and night shifts, the within-subgroup variation will be polluted, making the control limits too wide to detect real issues.

The subgroup size and sampling frequency should be clearly defined in the control plan and not changed arbitrarily. Any change in the subgroup size or sampling frequency invalidates the control limits.

Q12 | If an Automatic Inspection Device Measures Every Piece, Do We Still Need Subgroups?

Conclusion: Full inspection does not mean you can skip subgroups; in fact, it makes it even more important to clearly define how subgroups are formed — because you have too much data.

A common misconception is: since full inspection provides all the data, why do we still need subgroups? The reason is that the purpose of a control chart is not to plot all the data, but to determine process stability. Too much data means too much noise, making it difficult to see the process clearly.

First, distinguish between two things: full inspection data is used to determine product qualification, while control charts are used to determine process stability. The handling methods are different.

Three uses of full inspection data:

1. Aggregate by time — every 500 pieces or every half hour, forming a data point that represents the process level, not a pile of product data.
2. Pay attention to the division within and between subgroups — full inspection data can form large subgroups with minimal within-subgroup variation and narrow control limits, increasing sensitivity but also increasing false alarms.
3. Use the results of automatic judgment as data for the nonconforming units chart, and simultaneously use the defects chart to track the number of defects, viewing both lines together.

Two practical tips: Write the selected criteria into the control plan or work instruction, clearly specifying which criteria to use, and do not let different shifts decide on their own; clearly define the response actions after an alarm — who checks, how quickly they respond, and what the judgment criteria are. The most common issue in the industry is not the lack of rules, but the lack of action after an alarm.

The value of a control chart lies in its ability to alert you before the product becomes nonconforming, significantly reducing the cost of action compared to waiting for customer returns.

3. Process Capability: Cpk, Ppk, and Thresholds

Q3 | What is the Difference Between Cpk and Ppk? Why Do the Two Numbers Conflict?

Conclusion: The numerators of the two indices are the same, but the denominators differ. The numerator is the distance from the mean to the nearest specification limit, and the denominators are three times the standard deviation — but one uses within-subgroup standard deviation, and the other uses overall standard deviation.

There is a constant equation: Cpk ÷ Ppk = σoverall ÷ σwithin. This ratio itself is a diagnostic indicator.

The two standard deviations measure different things:

• Within-subgroup standard deviation only calculates the differences within the same subgroup, representing inherent variation under consistent conditions.
• Overall standard deviation calculates all individual values mixed together, equal to the within-subgroup variation plus the amount the process center has shifted between subgroups.

Therefore, the overall standard deviation is always ≥ the within-subgroup standard deviation, and Ppk is usually ≤ Cpk.

Why do the two numbers conflict? Because they answer different questions:

So, when the process engineer says, "Cpk is 1.35, the equipment is fine," and the quality engineer says, "Ppk is 0.93, it will exceed the limits," both are correct. The complete conclusion is: The equipment is sufficient, but the process center is drifting.

A related correction: many trainings say "Cpk is short-term, Ppk is long-term," which is not entirely wrong but can be misleading. It refers to the time scale of the variation sources, not the duration of data collection. Collecting data for 20 minutes and mixing it all together still results in Ppk.

The difference between the two is the amount of process center drift. The larger the difference, the more you should look for the action causing the drift.

Q4 | Is Cpk 1.33 Always Qualified? Where Does the Number 1.67 Come From?

Conclusion: 1.33 is not a legal requirement, but an industry-standard threshold. Meeting it does not guarantee qualification, and not meeting it does not necessarily mean disqualification.

This number is derived from the normal distribution:

The choice of threshold essentially depends on how much cost you are willing to incur to achieve a certain nonconformance rate.

Three essential limitations to remember:

1. This is a theoretical value — the process must be in control and the data must approximate a normal distribution. If the process is not in control, the calculated Cpk is merely decorative.
2. It only guarantees the predicted nonconformance rate, not the actual rate — measurement errors, sampling representativeness, and specification reasonableness can all cause deviations.
3. Customer requirements take precedence — follow the specific requirements in the customer's special requirements, regardless of industry practices.

The Cpk threshold is a level of nonconformance rate derived from the distribution, not a certificate of qualification. Before looking at it, ensure the process is stable, the measuring instruments are accurate, and the data is genuine.

Q6 | Can Cpk Be Calculated if the Data Is Not Normally Distributed?

Conclusion: It can be calculated, but the result may mislead you.

The derivation of Cpk is based on the normal distribution, assuming that three times the standard deviation is a fixed probability boundary. When the data is heavily skewed, this assumption no longer holds.

Three common pitfalls:

1. Characteristics with one-sided specifications — such as positional tolerances, roughness, and impurity content, which naturally cluster at zero and are heavily right-skewed.
2. Count data — nonconforming units and defects, which are inherently binomial or Poisson, not normal.
3. Characteristics with physical boundaries — hardness and thickness cannot be less than zero, and the distribution near the boundary will be skewed.

Four methods, ranked by cost:

1. First, plot the data to see the shape — histograms and probability plots can quickly show whether the data is skewed. This is almost cost-free, but many people skip this step.
2. Transform the data — the most common transformations are logarithmic and Box-Cox, which can bring the data back to a near-normal distribution.
3. Use non-normal capability analysis — most statistical software supports this, such as fitting with Weibull or Gamma distributions.
4. Use the percentage method — instead of Cpk, directly calculate the actual nonconformance rate and PPM, which does not rely on distribution assumptions.

The ability to calculate the formula does not mean the conclusion is reliable. First, examine the distribution shape, then decide which method to use. When the data is clearly non-normal, it is better to directly look at the PPM — Ppk is also more robust under non-normal conditions because it uses the entire data range.

4. What to Do When an Alarm Occurs

Q7 | The Control Chart Alarms, but the Product Re-tests合格. Should We Stop the Line?

Conclusion: Do not ignore the alarm just because the re-test is合格, nor should you stop the line immediately upon seeing an alarm. First, distinguish the level of the alarm.

First level, determine if the alarm is real or false. First, rule out measurement system and data entry issues — if the measuring instrument has drifted, the probe is dirty, or the data is entered incorrectly, the control chart may alarm even though the product is actually合格. The first step is always to verify the measurement, not to immediately stop the production line.

Second level, a real alarm indicates that the process has changed, but the product has not yet exceeded the limits. This is the most common state and the most valuable moment for a control chart. The process center is starting to drift, and the variation is increasing, but the magnitude is still small, and the product is within the specifications. At this point, stopping the machine for adjustment is the least costly; waiting until the product exceeds the limits will result in a batch of nonconforming products.

Third level, determine whether to stop based on two factors:

• Severity of the alarm — systematic signals such as multiple consecutive points on the same side or continuous increasing or decreasing trends should be addressed; isolated single points exceeding the limit can be observed for a while.
• Capability margin — if the Cpk is only 1.1, the margin is small, and the alarm should be acted upon immediately.

Four standard actions: verify the measurement to rule out data issues → isolate and clearly label the products from this period, do not mix batches → determine whether the cause is common or special, and address the root cause if it is special → after rectification, re-sample to verify and confirm that the process has returned to control before resuming production.

The value of a control chart lies in its ability to alert you before the product becomes nonconforming, significantly reducing the cost of action compared to waiting for customer returns.

Q8 | Are All Eight Out-of-Control Criteria Necessary? Which Ones Should Be Used in the Field?

Conclusion: Not all eight criteria are necessary; prioritize the use of three in the field.

The essence of the eight criteria is to find a balance between two types of errors: being too lenient and missing real anomalies, or being too strict and generating too many false alarms, leading operators to lose faith in the chart. A control chart that no one looks at is as good as none.

Prioritize these three criteria:

1. A single point exceeds three standard deviations — the most classic and least prone to false alarms, one criterion is enough to catch sudden major anomalies. Always enable this.
2. Nine consecutive points on the same side of the center line — this catches center shifts. A single point may not show it, but nine consecutive points on the same side have a probability of only about 0.4%, indicating a significant center shift.
3. Six consecutive points increasing or decreasing — this catches trends. Typical scenarios include tool wear, catalyst degradation, and filter clogging, where the issue gradually worsens, and individual points never exceed the limits, but the trend is clear.

These three criteria cover the majority of field issues: sudden, shifts, and trends. The other criteria are more useful for detailed analysis and automated monitoring; too many rules for manual chart reading can lead to operator fatigue.

Two practical suggestions: write the selected criteria into the control plan or work instruction, clearly specifying which criteria to use, and do not let different shifts decide on their own; clearly define the response actions after an alarm — who checks, how quickly they respond, and what the judgment criteria are. The most common issue in the industry is not the lack of rules, but the lack of action after an alarm.

The value of a control chart lies in its ability to alert you before the product becomes nonconforming, significantly reducing the cost of action compared to waiting for customer returns.

5. Count Data Control Charts

Q9 | Why Are Nonconforming Units Charts and Defects Charts Two Different Charts? Aren't They the Same Thing?

Conclusion: They are not the same thing. One counts units, the other counts defects.

• Nonconforming units count units — a product is either合格or非conforming, with no in-between. Therefore, the upper limit of nonconforming units is the sample size; if you sample 20 pieces, the maximum number of nonconforming units is 20.
• Defects count defects — a single product can have multiple issues, counted individually, so the number of defects can far exceed the sample size.

A specific example: if you sample 10 pieces and find 14 scratches and 9 pinholes, totaling 23 defects, but only 3 pieces are judged as nonconforming.

For the nonconforming units chart, the data point is 3
For the defects chart, the data point is 23
Per piece, it is 2.3 defects per piece

These three numbers describe completely different facts.

Why should both charts be used? Because they manage different aspects:

• Nonconforming units correspond to delivery risks and customer complaints — customers receive nonconforming units, not defects.
• Defects correspond to rework costs and internal losses — a single nonconforming unit with seven or eight defects will have rework time determined by the number of defects.

More critically, defects are leading indicators. When the process begins to deteriorate, defects often increase first, while nonconforming units have not yet appeared — because defects initially occur in non-critical areas or are still within the acceptance limits, the entire piece is still judged as合格. When the nonconformance rate starts to rise, the problem has already accumulated.

If the data is naturally countable by defect (scratches, pitting, porosity, pinholes), use the defects chart. Using the nonconforming units chart for such characteristics is like erasing information.

6. Chart Selection and Control Limit Maintenance

Q11 | When to Use the Mean and Range Chart vs. the Individual and Moving Range Chart?

Conclusion: Use the mean and range chart if you can take subgroups; use the individual and moving range chart if you can only take individual measurements.

The mean and range chart requires you to take subgroups — several pieces (three to five) taken continuously under the same conditions within the same time period. The advantage is that it provides within-subgroup information, allowing you to monitor both position and variation, with high sensitivity.

The individual and moving range chart is used when you can only take one measurement at a time: destructive testing can only destroy one piece at a time; if you only produce one batch per day, you have only one data point per day; if the testing cost is too high, you can only sample one.

Why can't you always use individual measurements? Because the individual chart has lower sensitivity. In the mean and range chart, the average of five pieces reduces noise, making the signal clearer; the individual chart lacks this smoothing, making small shifts difficult to detect, and is only suitable for monitoring larger shifts.

Three practical points:

1. Always use the individual and moving range chart together — using only the individual chart can lead to misjudgment.
2. The individual chart requires a larger data set — at least 20 points, ideally 25 to 30.
3. If the data has a clear trend or autocorrelation (such as temperature, pressure, and liquid level, which are real-time continuous variables), the individual chart will generate many false alarms — either difference the data first or switch to an EWMA or CUSUM chart.

Use the mean and range chart if you can take subgroups; use the individual chart as a fallback, and always use it with the moving range chart.

Q10 | Should Control Limits Be Recalculated After Process Improvements? How Often?

Conclusion: Recalculate if the process changes; do not recalculate unnecessarily if it has not changed.

Control limits are calculated from process data, describing the historical state of the process. If the process changes, the original control limits no longer describe the current process.

Four situations where recalculation is mandatory:

1. Process parameters have changed — temperature, pressure, speed, formula, etc. Any change in these parameters alters the variation structure.
2. Equipment or molds have been changed — a common scenario that many people overlook.
3. Key materials or suppliers have been changed
4. Major quality anomalies have occurred and been rectified — the process before and after rectification is not the same.

Situations where recalculation is not necessary: the process itself has not changed, and only a few points occasionally exceed the limits. Do not panic and recalculate — recalculation means absorbing the anomaly points, widening the control limits, and failing to detect real issues in the future. This is called using control limits to accommodate data, which is very dangerous. The correct approach is to first investigate the cause of the anomaly: if it is a one-time event due to a special cause, record, isolate, and analyze it, keeping the control limits unchanged.

Three principles for how often to recalculate:

1. Event-driven is preferred over time-driven — recalculate when the above four situations occur, and do not change if they do not.
2. Regularly review the control limits — check the control limits and data every quarter to see if they still match, but a review does not mean a mandatory recalculation.
3. Leave a trace when control limits are no longer used — record the old limits, the time period, and the reason for recalculation, as this evidence chain is important for audits.

Control limits are a mirror of the process. Change the mirror if the process changes; do not polish the mirror just to make the chart look good if the process has not changed.

7. When Data Is Insufficient

Q5 | Can a Control Chart Be Drawn with Only One or Two Batches of Data?

Conclusion: It can be drawn, but it cannot be used to determine process stability. Insufficient data makes the control limits unreliable.

Control charts have two purposes: determining stability and estimating variation. Insufficient data makes both purposes unachievable.

A common rule of thumb is to use around 25 subgroups. With fewer than 20 subgroups, the estimation error of the control limits is significant — because the control limits themselves are estimated, and a small sample size leads to a wide range of estimates. Using unreliable limits to judge anomalies will significantly increase both false alarms and missed alarms.

Four approaches, ranked by priority, when data is insufficient:

1. Extend the data collection period — gather more data over several weeks to form enough subgroups, the most stable approach.
2. For processes with few batches and long cycles, use the individual and moving range chart — a single data point can still be plotted, but you need at least 20 points, ideally 25 to 30.
3. If even this is not possible (e.g., a product is only produced twice a year), do not force a control chart — use a pre-control chart or first article inspection plus batch trend tracking. Using the wrong tool is more dangerous than using no tool.
4. For short-term projects, combine data from similar processes or use engineering judgment to set temporary control limits, replacing them when enough data is available.

One final reminder: clearly label the data used to establish the control limits. It is not embarrassing to say, "These are temporary limits" when data is insufficient; using temporary limits as permanent ones for three years without change is the real problem.

When data is insufficient, it is better to use other tools rather than rely on unreliable limits to convince yourself that the process is fine.

Appendix: Quick Reference Table for the 12 Questions

The 12 most frequent SPC questions boil down to one key point: first, distinguish between "how the process is fluctuating" and "what the product requirements are," and the rest of the discussion will have a solid foundation.

Knowledge code: 6.3.1

Version: v20261017

Author: QTank QTank is dedicated to providing systematic professional knowledge, methodologies, and practical tools for quality management practitioners, helping companies continuously improve their quality capabilities.