SPC for Complex Process Environments

Introduction and Context

Statistical process control (SPC) was initially developed for relatively stable, repetitive production: one product, one process, frequent observations, and approximately independent measurements. Contemporary operations rarely satisfy all these conditions. Pharmaceutical plants manufacture multiple strengths on shared equipment; semiconductor facilities monitor correlated characteristics across many process stages; hospitals and service organizations generate irregular, time-dependent data; and high-mix factories may produce only a few units of each configuration before changing over.

In such environments, conventional Shewhart, cumulative sum (CUSUM), or exponentially weighted moving average (EWMA) charts may give misleading signals. A chart can signal because the product changed, not because the process deteriorated; autocorrelation can inflate apparent evidence of instability; changing means and variances can make fixed limits obsolete; and a normal-theory chart can be inappropriate for skewed, bounded, count, or heavily tailed data.

Complex-environment SPC therefore requires a design logic rather than a single chart recipe. The practitioner must identify the observational unit, separate product and process effects, model dependence and non-stationarity, select an appropriate statistic, and connect signals to technically credible action. This chapter develops that logic through short-run and standardized charts, mixed-production monitoring, batch and profile methods, multivariate methods, autocorrelation, non-normal data, and implementation governance.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Explain why conventional control charts can fail in high-mix, batch, autocorrelated, and non-stationary processes.
2. Design short-run and standardized control charts for mixed products and limited production runs.
3. Distinguish common-cause variation from product, batch, shift, and environmental effects.
4. Apply multivariate monitoring using covariance-aware statistics such as Hotelling’s T².
5. Select and interpret profile-monitoring methods for functional or curve-shaped relationships.
6. Diagnose autocorrelation and non-stationarity before setting control limits.
7. Select transformations, distribution-specific charts, or model-based methods for non-normal data.
8. Integrate advanced SPC with measurement systems, process engineering, escalation, and governance.

1. A Design Model for Complex SPC

A useful starting point is to represent an observation as:

Here, y_i,t is the measured response for characteristic i at time t; x_i,t represents controllable process inputs; p_i identifies product or recipe; b_t identifies batch, lot, campaign, or equipment state; e_t represents environmental or operational conditions; and ε_i,t is unexplained measurement and process variation. The expression is conceptual rather than a claim that every process is linear.

The central question is not simply “Is the point outside the limits?” It is: “Has the conditional distribution of the response changed, after accounting for the product, batch, operating regime, and dependence structure?” This distinction separates advanced SPC from mechanical charting.

1.1 Sources of complexity

1.2 Stability, capability, and conformance

SPC primarily addresses stability: whether the process distribution is predictable over time. Capability compares a stable process distribution with specification limits. These are different questions. A process can be stable but incapable, or apparently capable during a selected period but unstable overall. In a mixed process, capability must be evaluated conditionally on product, recipe, or operating regime; pooling unlike populations can produce an artificial mean and variance.

2. Short-Run and Standardized SPC

2.1 Why ordinary limits fail in short runs

A traditional X̄ chart assumes rational subgroups from a sufficiently homogeneous process and estimates the within-subgroup standard deviation. In short-run production, each subgroup may contain different products or there may be too few subgroups to estimate variation reliably. A common but weak response is to combine unlike observations and calculate one mean and standard deviation. This confounds product differences with process variation.

Short-run SPC instead uses a common scale or a reference distribution. The reference may come from validated historical data, a pooled within-product estimate, a designed experiment, or a hierarchical model. The reference must be reviewed whenever materials, equipment, measurement methods, or specifications change.

2.2 Standardization

For product or characteristic j, a standardized observation is commonly defined as:

where μ_j is the accepted in-control center and σ_j is the corresponding in-control standard deviation. The numerator and denominator use the same measurement units, so z_j,t is dimensionless. If product-specific specifications are central to the decision, a specification-based standardization may be used, but specification limits must not be confused with statistically estimated control limits.

For a subgroup mean of size n, assuming independent observations with common variance σ_j²,

The standard error σ_j/√n is valid only when the subgroup observations are appropriately comparable and independence, or an appropriate effective variance model, is reasonable.

A standardized chart may use center 0 and limits ±3 for a Shewhart-style design. The “three-sigma” convention is a decision rule, not a universal law. It should be modified when false-alarm costs, sampling frequency, distributional shape, or multiple testing make the nominal signal rate unsuitable.

2.3 Short-run chart design

A defensible procedure is:

1. Define the process family and determine which products are genuinely comparable.
2. Establish product-specific centers and within-product variation.
3. Estimate variation from rationally formed, stable data; do not use specification width as a substitute for process variation.
4. Standardize observations or use a regression/model-based residual.
5. Plot the standardized statistic in chronological order while retaining product identity.
6. Apply supplementary run rules only after considering their false-alarm effect.
7. Investigate signals in the original engineering units.

A robust pooled estimate can be preferable to a noisy product-specific estimate when some products have few observations. However, pooling is justified only when a common variation assumption is supported by process knowledge and diagnostics.

Worked example: standardized short-run monitoring

An assembly process produces three product variants. For variant A, the validated center is 50.0 mm and the within-product standard deviation is 0.20 mm. A subgroup of four units has mean 50.36 mm. The standardized subgroup value is:

The point exceeds a conventional upper boundary of 3. It is a signal for investigation, not automatic proof of a special cause. The investigator should check measurement status, tool adjustment, material identity, and the four original readings. If the estimated variation is itself uncertain, the signal should be classified with that uncertainty documented.

2.4 Standardization is not normalization of reality

A standardized chart can make unlike products comparable, but it does not eliminate confounding. A product may have a genuinely different response to temperature, a different measurement error, or a different failure mechanism. Standardized monitoring should therefore display product, recipe, equipment, and operator identifiers, and should be supplemented by stratified analysis when signals cluster within one category.

3. Mixed Production, Batches, and High-Mix Operations

3.1 Hierarchical variation

Batch processes commonly contain at least three variance components:

where α_b is a batch effect and ε_bt is within-batch variation. Under the usual random-effects assumptions, α_b has variance σ_b², ε_bt has variance σ_w², and the variance of an individual observation is σ_b² + σ_w². The assumptions include appropriate independence between batch effects and within-batch errors and a stable variance structure. These assumptions must be checked rather than presumed.

A batch-level chart may monitor batch means, while a within-batch chart monitors dispersion or residuals. Charting only individual observations can conceal a systematic batch shift; charting only batch means can conceal within-batch deterioration.

3.2 Mixed-model and residual monitoring

When product and operating variables explain predictable differences, fit a process model such as:

Here, β coefficients describe fixed effects, u_p(t) may represent product effects, v_b(t) batch effects, and ε_t the residual. Monitoring standardized residuals can focus SPC on unexplained process behavior. The model must be validated out of sample and periodically reviewed; otherwise, an incorrect model can hide a real shift.

3.3 High-mix operating rules

High-mix facilities benefit from a layered architecture:

• Recipe-level checks: confirm set points, tooling, material, and specification identity.
• Family-level standardized chart: monitor comparable products on a common scale.
• Equipment-health chart: monitor variables common to all products, such as pressure or spindle load.
• Batch and campaign review: distinguish startup, steady-state, changeover, and shutdown behavior.
• Exception workflow: require a documented disposition for every statistically meaningful signal.

Changeovers should not be silently treated as ordinary observations. A planned transient may require a separate phase or a startup chart; an unplanned extension of the transient is a process signal.

4. Multivariate SPC

4.1 Why separate charts are insufficient

Suppose diameter, roundness, and surface roughness are measured together. Separate charts may each remain within limits while the joint combination is unusual. Conversely, correlated characteristics can produce excessive alarms when each is charted independently. Multivariate methods account for covariance and evaluate the joint state.

For a vector x of p characteristics, with in-control mean vector μ and covariance matrix Σ, the Mahalanobis distance is:

Σ must be positive definite and estimated from representative in-control data. The result is dimensionless and accounts for scale and correlation. If variables are strongly collinear, regularization, dimension reduction, or removal of redundant measures may be necessary.

4.2 Hotelling’s T² chart

For an individual observation when parameters are treated as known, a common statistic is:

This is the same quadratic form as D², often interpreted in a control-chart setting. For a subgroup mean ̄x of size n,

These formulas describe the statistic, not a universal control limit. Control limits depend on whether μ and Σ are known or estimated, the subgroup size, dimension, phase of monitoring, and distributional assumptions. Under multivariate normality and known parameters, T² for an individual follows a chi-square distribution with p degrees of freedom. With estimated parameters, finite-sample adjustments based on the F distribution are required, and the exact form depends on the sampling design.

4.3 Diagnosis after a multivariate signal

A multivariate signal identifies an unusual joint observation but does not automatically identify the responsible characteristic. Useful follow-up methods include contribution plots, conditional analysis, parallel univariate charts, engineering checks, and comparison with historical failure signatures. Contributions can be unstable when variables are highly correlated, so they should be treated as diagnostic aids rather than causal proof.

4.4 Modern multivariate extensions

When p is large relative to the reference sample, covariance estimation becomes unstable. Principal component monitoring can represent dominant covariance directions; partial least squares may be useful when quality variables are related to process predictors; and shrinkage covariance estimators can improve conditioning. These methods introduce tuning choices and potential information loss. They require locked validation data and transparent model governance.

5. Profile Monitoring

A profile is a response measured as a function of a continuous or categorical explanatory variable. Examples include temperature versus time in a thermal cycle, thickness across a web, or concentration versus wavelength. Monitoring only the average response can miss a slope, curvature, or localized shape change.

A linear profile may be written as:

where profile i is observed at locations x_j, β_0i is its intercept, β_1i its slope, and ε_ij error. The vector of estimated coefficients (β̂_0i, β̂_1i) can be monitored jointly. For a nonlinear profile, polynomial, spline, mechanistic, or functional principal-component representations may be appropriate, but complexity should be justified by prediction and diagnosis, not visual appeal.

A profile model must address:

• common versus varying measurement locations;
• correlation among points within a profile;
• heteroscedasticity along the profile;
• missing or irregularly sampled locations;
• alignment or phase variation;
• whether the profile is a response curve or a repeated time series.

A profile signal should be translated back into an engineering interpretation: changed gain, offset, time constant, peak location, area, or local defect. This makes the chart actionable.

6. Autocorrelation and Non-Stationarity

6.1 Autocorrelation

Autocorrelation means that observations close in time are statistically related. For a stationary series with mean μ and variance σ², lag-k autocorrelation is:

A positive ρ_k reduces the effective information in a sequence and can cause ordinary limits to signal too frequently. The problem is not solved merely by widening limits, because the chart may then miss meaningful changes.

One response is to model the expected process dynamics and monitor residuals. For a first-order autoregressive model:

where c is a constant, φ is the autoregressive coefficient, and ε_t is an innovation with mean zero. Stationarity for this simple AR(1) model requires |φ| < 1. If the model is adequate, residuals should have little remaining autocorrelation and can be charted. The model is not a license to remove real assignable variation: unusual residuals and unusual parameter behavior both require review.

Other responses include sampling less frequently, using subgroup summaries, applying time-series-aware limits, or using dynamic CUSUM/EWMA methods. The choice depends on whether the dependence is a physical feature to be modeled or a sampling artifact that can be redesigned.

6.2 Non-stationarity

A process is non-stationary when its probabilistic behavior changes over time, such as through drift, changing variance, seasonality, wear, or regime shifts. A rolling average can adapt to drift but may absorb a special cause into the baseline. Adaptive limits therefore require explicit rules for reference-window length, exclusion of abnormal periods, reset authority, and retrospective review.

Useful diagnostics include time plots, rolling center and spread, autocorrelation functions, change-point methods, stratification by phase, and comparison of pre- and post-change distributions. Non-stationarity should be explained physically whenever possible. A control chart should not be used to conceal a process that is inherently unstable or poorly specified.

7. Non-Normal, Discrete, and Bounded Data

The normal distribution is convenient but not universal. Defect counts, durations, waiting times, proportions, concentrations, and fatigue-life measurements often have skewness, discreteness, censoring, or natural bounds.

Approaches include:

1. Transformation: A logarithm can reduce right skew for positive data; a square-root transformation is often useful for counts. The transformation must be defined, interpreted, and validated.
2. Distribution-specific charts: Poisson-based charts suit counts under appropriate assumptions; binomial charts suit defectives or proportions with known denominators; time-between-event charts may suit rare events.
3. Quantile or empirical limits: Useful when sufficient stable historical data exist, but sensitive to sample size and baseline contamination.
4. Generalized linear or distributional models: Model the conditional mean and variance using predictors and an appropriate link or distribution.
5. Robust methods: Median, trimmed statistics, robust covariance, or rank-based methods can reduce sensitivity to extreme observations, but may reduce sensitivity to specific failure modes.

For a binomial proportion p̂ = d/n, where d is the number of defectives in n independent trials, the approximate standard deviation is √(p(1−p)/n). This approximation is appropriate only when the binomial assumptions and a suitable baseline p are reasonable. Small counts, changing denominators, overdispersion, or dependence require exact, beta-binomial, or other models.

For count data, a Poisson model assumes equal mean and variance conditional on the modeled factors. Overdispersion indicates that this assumption is inadequate; blindly using Poisson limits can create misleading alarms. A negative-binomial or hierarchical model may be more suitable.

8. A Unified Implementation Framework

Advanced SPC should be implemented as a controlled measurement-and-decision system.

Step 1: Define the decision

Specify what action a signal should trigger: stop, quarantine, adjust, inspect, sample more intensively, or merely review. A chart without an action rule creates noise rather than control.

Step 2: Define the observational unit

Clarify whether the unit is an individual, subgroup, batch, profile, campaign, shift, or time interval. Record product and process context with each observation.

Step 3: Validate measurement capability

Gauge bias, repeatability, reproducibility, resolution, calibration, and data integrity. A sophisticated chart cannot compensate for an unstable measurement system.

Step 4: Establish a phase-I baseline

Use historical data to identify special causes, estimate centers and covariance, test model assumptions, and document exclusions. Phase I is exploratory and retrospective; Phase II is prospective monitoring.

Step 5: Choose the monitoring statistic

Select standardized residuals, batch summaries, multivariate statistics, profiles, time-series residuals, or distribution-specific statistics according to the data-generating process.

Step 6: Validate alarm performance

Assess false-alarm probability, detection delay, robustness to parameter estimation, missing data, and multiple characteristics. Simulation is often more informative than relying solely on asymptotic limits.

Step 7: Link signals to investigation

Use a reaction plan with ownership, containment, confirmation tests, root-cause analysis, correction, and effectiveness review. Preserve the original data; do not overwrite a signal by recalculating limits without authorization.

Step 8: Govern model changes

Version control baselines, transformations, features, covariance matrices, software, and limits. Define who may approve a reset and how prior performance is retained for audit and learning.

Illustrative Case: High-Mix Thermal Treatment Line

Background

A fictional manufacturer heat-treats six alloy families in two furnaces. Each campaign lasts between two and eight hours, and product changes occur several times per day. Quality engineers measure hardness, distortion, and residual stress. Furnace temperature is recorded every minute. The original practice was to place all hardness values on one X̄ chart.

Problem

The chart generated frequent alarms during product changes but failed to highlight a gradual furnace-control problem. Investigation showed that alloy families had different nominal hardness, residual stress was correlated with hardness, and temperature readings were strongly autocorrelated. Campaign-level effects were also present.

Analysis

The team first separated recipe identity, furnace, campaign, and startup status. A measurement study found acceptable repeatability but a small calibration bias, which was corrected. Product-specific centers and within-family variation were estimated from stable campaigns. A mixed model explained hardness using alloy family and measured temperature, with campaign as a random effect. Residual diagnostics showed remaining serial dependence in minute-by-minute temperature data.

For quality characteristics, standardized residuals were monitored by alloy family, and hardness and residual stress were also monitored jointly using a covariance matrix estimated from in-control campaigns. For temperature, an autoregressive model was fitted and innovations were charted rather than raw one-minute readings. Startup observations were placed in a separate phase because the physical warm-up transient was expected.

Approach and solution

The operating system used three levels:

• recipe verification and startup checks;
• campaign-level standardized and multivariate monitoring;
• furnace-health monitoring through time-series residuals.

A signal required the operator to record material, recipe, furnace, time, and recent adjustments before any limit reset. A multivariate alarm prompted review of both hardness and residual stress contributions, while a temperature-residual alarm prompted examination of sensors and control valves.

Outcome

The number of product-change alarms declined because unlike alloys were no longer pooled. A developing control-valve problem was detected through sustained temperature innovations before a large quality shift occurred. The organization retained the old chart as a limited conformance display but no longer treated it as the primary process-control chart.

Learning points

The case demonstrates that advanced SPC is an architecture, not merely a more complicated limit calculation. Product context, hierarchy, dependence, and action design were as important as the statistical statistics themselves. It also illustrates a governance principle: models should focus attention, but engineering verification remains necessary before disposition.

Technology, AI, and Governance

Modern manufacturing execution systems and historians can automatically capture recipe, equipment, environmental, and quality data. Machine learning can support anomaly detection, representation learning, and nonlinear prediction, especially when profiles or high-dimensional sensor streams are involved. However, an anomaly score is not automatically an SPC signal. AI models may drift, encode historical operator decisions, obscure causal interpretation, and produce unstable thresholds under changing product mix.

A responsible deployment should include data lineage, human-readable features, separate training and validation periods, drift monitoring, access controls, cybersecurity, and a defined human decision authority. Regulated industries should preserve electronic records, approval histories, audit trails, and validated software according to applicable quality-system requirements. Privacy and fairness also matter when worker, patient, or customer data are used as process predictors.

Challenges and Limitations

Advanced methods do not remove the fundamental limitations of observational data. A poor baseline contaminates every limit. A covariance matrix estimated from too few observations can be nearly singular. Overfitted regression residuals may appear stable because the model has memorized the data. Adaptive limits can normalize deterioration. Multiple charts and repeated testing increase the probability of false alarms. Conversely, excessive stratification can reduce sensitivity by fragmenting the data.

The remedy is disciplined simplicity: use the least complex method that represents the process adequately, validate assumptions, preserve raw and contextual data, and test performance under plausible disturbances. Statistical evidence should be combined with engineering knowledge, measurement verification, and risk-based decision rules.

Best-Practice Checklist

• Define the process family and observational unit before selecting a chart.
• Separate specifications from control limits.
• Estimate variation within rationally comparable groups.
• Retain product, batch, phase, equipment, and operator identifiers.
• Diagnose dependence, drift, changing variance, and distributional form.
• Use multivariate methods when joint behavior matters, not merely because many variables exist.
• Monitor profile parameters or residual functions when shape is important.
• Validate covariance and time-series models on data not used to fit them.
• Quantify false alarms and detection delay through analytical or simulation studies.
• Connect every signal to a documented, risk-based reaction plan.
• Govern model, baseline, and limit changes with version control and authorization.

Chapter Summary

Complex process environments require SPC designs that respect product mix, batch hierarchy, correlated characteristics, profile shape, temporal dependence, non-stationarity, and non-normal distributions. Short-run and standardized charts create comparability only when reference centers and variation estimates are meaningful. Batch and mixed-model methods separate within- and between-group effects. Multivariate statistics such as Mahalanobis distance and Hotelling’s T² evaluate joint behavior, while profile monitoring detects changes in intercept, slope, curvature, or other functional features. Autocorrelation should be modeled or addressed through sampling design rather than ignored. Non-normal data require transformations, distribution-specific methods, robust procedures, or generalized models. Successful implementation combines statistical validity with measurement assurance, engineering interpretation, action planning, software governance, and ethical control of automated decisions.

Key Terms

Autocorrelation: Statistical dependence between observations separated by a time lag.

Batch effect: Systematic variation associated with a batch, lot, campaign, or production run.

Baseline: Reference data and parameter estimates used to define in-control behavior.

Change point: A time or condition at which the process distribution changes.

Covariance matrix: Matrix describing variances and pairwise covariances among characteristics.

High-mix production: Operations producing many products or configurations in relatively small quantities.

Hotelling’s T²: A covariance-aware multivariate distance used for monitoring vectors or subgroup means.

Mahalanobis distance: Standardized multivariate distance accounting for scale and correlation.

Non-stationarity: Change over time in a process mean, variance, dependence, or distribution.

Profile: A response observed as a function of an explanatory variable or location.

Rational subgroup: A group formed so that within-group variation represents a meaningful source of routine process variation.

Residual: Difference between an observed value and a model-predicted value.

Short-run SPC: Control-chart practice for limited, mixed, or intermittent production runs.

Standardization: Rescaling an observation relative to a reference center and standard deviation.

Phase I: Retrospective analysis used to establish and clean an SPC baseline.

Phase II: Prospective monitoring using a previously established baseline.

Overdispersion: Variability greater than that predicted by a reference count or binomial model.

Reaction plan: Documented sequence of actions following a control or diagnostic signal.

Review Questions

Short Answer Questions

1. Why can pooling different products on one conventional control chart create misleading signals?
2. Define standardization and explain why the standardized value is dimensionless.
3. What information does a multivariate statistic add beyond separate univariate charts?
4. Why can autocorrelation cause ordinary control limits to signal too frequently?
5. Distinguish Phase I analysis from Phase II monitoring.

Long Answer Questions

1. Explain how you would design an SPC system for a high-mix batch process, including data structure, baseline estimation, chart selection, and reaction planning.
2. Discuss the assumptions, interpretation, and limitations of Hotelling’s T² chart.
3. Compare transformation, distribution-specific charts, robust methods, and generalized models for non-normal data.
4. Explain how profile monitoring differs from monitoring a single summary statistic, using an engineering example.
5. Develop a governance framework for adaptive or AI-assisted SPC in a regulated manufacturing environment.

Analytical / Application Questions

1. A process has three products with different means but similar within-product standard deviations. Propose a standardized short-run chart design and identify the evidence needed before pooling the variation estimates.
2. A time-series chart shows repeated alarms, but the raw observations have strong positive lag-one correlation. Explain the diagnostic steps and recommend two technically defensible remedies.
3. Two quality characteristics are individually within limits, yet their covariance-aware statistic signals. Describe a structured investigation that distinguishes a real joint shift from an unstable covariance estimate.

Multiple Choice Questions with Answers

Question 1

Which feature most directly motivates the use of standardized short-run SPC?

A. A single product produced continuously for months

B. Many products with different centers produced in limited runs

C. A measurement system with no recorded observations

D. A process with specifications but no process data

Correct Answer: B. Many products with different centers produced in limited runs

Question 2

In the expression z = (x − μ)/σ, what does μ represent?

A. The specification width

B. The sample size

C. The accepted in-control center

D. The covariance between variables

Correct Answer: C. The accepted in-control center

Question 3

What is the principal purpose of the covariance matrix in Mahalanobis distance?

A. To remove all measurement error

B. To account for scale and correlation among characteristics

C. To convert counts into proportions

D. To identify the production batch automatically

Correct Answer: B. To account for scale and correlation among characteristics

Question 4

For the simple AR(1) model Y_t = c + φY_t−1 + ε_t, which condition is associated with stationarity?

A. φ = 0 only

B. |φ| < 1

C. |φ| > 1

D. φ must equal the process mean

Correct Answer: B. |φ| < 1

Question 5

What does a profile-monitoring method primarily detect?

A. Only the total number of units produced

B. Changes in the shape or relationship of a response curve

C. Only changes in specification limits

D. The identity of the operator

Correct Answer: B. Changes in the shape or relationship of a response curve

Question 6

A batch model is written as Y_bt = μ + α_b + ε_bt. What does α_b represent?

A. A batch-specific effect

B. The measurement unit

C. The subgroup sample size

D. The specification limit

Correct Answer: A. A batch-specific effect

Question 7

Why can an ordinary chart applied to autocorrelated observations produce excessive alarms?

A. Dependence reduces the apparent number of observations and changes signal behavior

B. Autocorrelation guarantees a normal distribution

C. Autocorrelation makes all observations independent

D. The control limits become specification limits

Correct Answer: A. Dependence reduces the apparent number of observations and changes signal behavior

Question 8

Which situation most strongly suggests a distribution-specific count chart rather than a normal-theory chart?

A. Continuous measurements with stable symmetric variation

B. Defect counts with an appropriate count-data model

C. A profile measured at many locations

D. A vector of highly correlated continuous characteristics

Correct Answer: B. Defect counts with an appropriate count-data model

Question 9

What is the main danger of resetting adaptive control limits whenever the process drifts?

A. The chart may absorb deterioration into the baseline

B. The measurement system becomes perfectly accurate

C. Product identity is automatically removed

D. The process becomes multivariate

Correct Answer: A. The chart may absorb deterioration into the baseline

Question 10

After a multivariate signal, what is the most appropriate immediate interpretation?

A. One specific variable is proven to be the root cause

B. The joint observation is unusual and requires follow-up diagnosis

C. All specifications have necessarily been violated

D. The covariance matrix is certainly correct

Correct Answer: B. The joint observation is unusual and requires follow-up diagnosis