6.2 Data Types & Classification

Data Types & Classification

Quick Answer: Six Sigma divides data into qualitative (categorical/attribute) and quantitative (numerical) forms, further distinguishing between discrete data (counted in distinct integer units) and continuous data (measured along an unbroken, infinitely divisible continuum). Under Stevens' measurement hierarchy, data is categorized into four progressive levels: Nominal (labels with no order), Ordinal (ranked categories with unequal intervals), Interval (equal intervals with an arbitrary zero), and Ratio (equal intervals with a true absolute zero). Continuous ratio data yields the highest statistical power with smaller sample sizes. Independent CSSYB study guide by OpenExamPrep.

Overview: Qualitative vs. Quantitative Data

Before collecting data in the Measure phase, a Six Sigma team must understand the nature and structure of the data generated by the process. Selecting the correct control chart, capability index, or hypothesis test depends entirely upon data classification. At the highest level, data is split into two domains:

1. Qualitative (Attribute) Data: Descriptive, non-numerical information that categorizes items based on characteristics, properties, or subjective attributes. Examples include customer gender, machine operator name, product color, failure category, or binary compliance (conforming vs. non-conforming). Qualitative data is inherently descriptive, requiring tallying or frequency counting before statistical tools can be applied.
2. Quantitative (Variable) Data: Objective numerical measurements or counts that describe how much, how many, or how long. Quantitative data inherently possesses numerical meaning, permitting arithmetic operations such as addition, averaging, and dispersion modeling. Examples include part diameter in millimeters, ambient oven temperature, transaction processing time in seconds, or annual scrap expense.

In DMAIC execution, teams frequently capture qualitative customer feedback during the Define phase (Voice of the Customer) and systematically translate it into quantitative, measurable Critical-to-Quality (CTQ) process metrics during the Measure phase.

---

Continuous (Variable) Data vs. Discrete (Attribute) Data

Within quantitative data, Six Sigma makes a crucial distinction between continuous and discrete data structures, which fundamentally governs sampling requirements and analytical precision.

Continuous (Variable) Data

Continuous data consists of numerical values measured along an unbroken, continuous scale. Between any two points on a continuous scale, an infinite number of intermediate values can exist, limited only by the precision and resolution of the measurement instrument.

• Physical Dimensions: Length, thickness, outer diameter, surface roughness, coating weight.
• Temporal Metrics: Total cycle time, customer queue wait time, machine downtime duration.
• Environmental & Operational Conditions: Temperature, hydraulic pressure, electrical voltage, fluid flow rate, torque.

The Statistical Advantage of Continuous Data

Continuous data provides the greatest depth of statistical information per observation. Because each measurement captures exact magnitude, distance from target, and fine gradations of variation, continuous data possesses high statistical power. Consequently, a Six Sigma team can evaluate process stability, estimate capability ($C_p, C_{pk}$), and detect subtle process mean shifts using relatively small sample sizes (typically $n = 30$ to $50$ observations).

Discrete (Attribute) Data

Discrete data consists of numerical values counted in distinct, separate whole-integer units or categorical buckets. Discrete data cannot be meaningfully divided into infinite fractions; there are no intermediate states between adjacent integers (e.g., an office may file 4 or 5 incorrect tax returns, but never 4.37 incorrect returns).

Discrete data generally falls into three statistical categories:

1. Binomial (Binary / Dichotomous) Data: Classifies units or events into one of two mutually exclusive states. Examples include pass/fail inspection, go/no-go gauging, on-time vs. late delivery, and paid vs. unpaid invoices. Binomial data forms the basis for $p$-charts (proportion non-conforming) and $np$-charts (number non-conforming).
2. Poisson (Count / Rate) Data: Counts the number of discrete defects or non-conformities occurring over a defined continuous area of opportunity (such as time, area, length, or volume). Examples include the number of paint blemishes per vehicle door, soldering flaws per printed circuit board, or customer billing complaints logged per month. Poisson data forms the basis for $c$-charts (count of defects per unit) and $u$-charts (defects per unit across varying inspection areas).
3. Multinomial Data: Observations classified into three or more distinct qualitative categories, such as sorting product defects into aesthetic, mechanical, electrical, or packaging failures.

Sample Size Penalties of Discrete Data

Discrete attribute data conveys significantly less information per observation than continuous variable data. Knowing that a part merely "passed" or "failed" provides no insight into whether the part was comfortably within tolerances or barely scraped past the specification limit. Because attribute data loses this fine detail, detecting process shifts or establishing high-confidence capability requires substantially larger sample sizes—often hundreds or thousands of observations ($n = 100$ to $1,000+$).

---

Stevens' Four Levels of Measurement (NOIR)

In 1946, Harvard psychologist Stanley Smith Stevens published a landmark paper in Science introducing a hierarchical taxonomy of measurement scales: Nominal, Ordinal, Interval, and Ratio (often memorized using the acronym NOIR). This hierarchy is cumulative: each successive level possesses all mathematical properties of the preceding levels, plus new capabilities.

Nominal (Lowest Level: Categorical Labels)
  └── Ordinal (Ranked Order, Unequal Spacing)
        └── Interval (Equal Intervals, Arbitrary Zero)
              └── Ratio (Highest Level: Absolute Zero, Full Math)

1. Nominal Scale (Labels & Categories)

The nominal scale represents the most basic level of measurement. Numbers or words serve strictly as qualitative labels or category identifiers with no inherent mathematical value, magnitude, or natural ordering.

• Examples: Machine identification codes (Press A, Press B, Press C), operator shift labels (Shift 1, Shift 2, Shift 3), failure modes (scratch, dent, porosity), zip codes, telephone area codes.
• Valid Mathematical Operations: Equivalence testing only ($=$ or $\neq$). Arithmetic operations such as addition, subtraction, multiplication, or calculating a mean are completely meaningless (e.g., averaging Zip Code 90210 and 10001 yields nonsense).
• Allowable Statistics: Frequency counts, percentages, proportion calculations, the Mode, and contingency table tests (such as Chi-Square $\chi^2$).

2. Ordinal Scale (Ranked Order)

The ordinal scale categorizes observations into a natural, sequential ranking or relative order. It indicates that one item has more or less of an attribute than another, but the distance or interval between adjacent ranks is unknown, unequal, or subjective.

• Examples: Customer satisfaction surveys utilizing Likert scales (1 = Very Dissatisfied, 2 = Dissatisfied, 3 = Neutral, 4 = Satisfied, 5 = Very Satisfied); job evaluation tiers (Entry, Mid, Senior, Executive); finish rankings in a race (1st, 2nd, 3rd place).
• Operational Nuance: In a 5-point customer satisfaction survey, the numerical difference between "Very Dissatisfied" (1) and "Dissatisfied" (2) is not mathematically identical to the psychological difference between "Satisfied" (4) and "Very Satisfied" (5). Similarly, the 1st place runner might beat 2nd place by 0.1 seconds, while 2nd place beats 3rd place by 12 seconds.
• Valid Mathematical Operations: Order comparisons ($>$ or $<$). Addition and subtraction cannot be legitimately performed on pure ordinal ranks.
• Allowable Statistics: Median, percentiles, quartiles, range of ranks, Spearman's rank correlation coefficient ($\rho$), and non-parametric hypothesis tests (e.g., Mann-Whitney $U$, Kruskal-Wallis). Reporting the arithmetic mean of Likert survey scores is a common operational error in business; the median provides the methodologically sound measure of central location.

3. Interval Scale (Equal Intervals Without Absolute Zero)

The interval scale provides quantitative measurements where the spacing or intervals between consecutive numbers are equal and constant across the entire scale. However, an interval scale lacks a true, non-arbitrary absolute zero point. Zero on an interval scale is simply an agreed-upon reference point, not the complete absence of the measured phenomenon.

• Examples: Temperature measured in degrees Celsius or Fahrenheit ($0^\circ\text{C}$ is the freezing point of water, not the total absence of heat energy); calendar years (the year 2026 does not mean 2,026 years since the dawn of time); standardized intelligence (IQ) scores.
• Mathematical Properties: Subtraction and addition are fully valid. The temperature difference between $20^\circ\text{C}$ and $30^\circ\text{C}$ ($10^\circ\text{C}$) is physically identical to the difference between $70^\circ\text{C}$ and $80^\circ\text{C}$ ($10^\circ\text{C}$). However, multiplication and division of absolute scale values are invalid because zero is arbitrary: $40^\circ\text{C}$ is not "twice as hot" as $20^\circ\text{C}$ (converting to the Kelvin scale reveals that $20^\circ\text{C} = 293.15\text{ K}$ and $40^\circ\text{C} = 313.15\text{ K}$, representing only a $6.8\%$ increase in thermal energy).
• Allowable Statistics: Arithmetic mean, standard deviation, variance, Pearson product-moment correlation ($r$), two-sample $t$-tests, and ANOVA. Ratios are invalid.

4. Ratio Scale (Equal Intervals With Absolute Zero)

The ratio scale represents the highest and most powerful level of measurement. It possesses all properties of the interval scale—equal intervals, clear order, and quantitative magnitude—with the crucial addition of a true, non-arbitrary absolute zero point. Zero on a ratio scale signifies the complete and total absence of the measured property.

• Examples: Cycle time in seconds (0 seconds indicates instantaneous completion), physical dimensions (weight, length, width, volume), financial metrics (cost of poor quality, revenue, scrap expense), electrical resistance (ohms), defect counts, temperature in Kelvin.
• Mathematical Properties: All mathematical operations are permissible: addition, subtraction, multiplication, division, and the calculation of meaningful ratios. An order that requires 60 seconds to process takes exactly twice as long as an order requiring 30 seconds ($60 / 30 = 2.0$).
• Allowable Statistics: All parametric and non-parametric statistical methods, including the geometric mean, harmonic mean, coefficient of variation ($CV$), linear regression, and advanced process capability indices.

---

Measurement Scales Summary Matrix

---

Practical Yellow Belt Strategy: Converting Attribute Data to Continuous Data

Because continuous ratio data provides superior statistical power with smaller sample sizes, Six Sigma teams should always endeavor to convert attribute measurements into continuous variables at the point of data collection:

• Poor Practice (Discrete Attribute): Inspecting machined shafts with a go/no-go ring gauge and simply recording parts as "Conforming" or "Non-conforming."
• Best Practice (Continuous Variable): Measuring the exact outer shaft diameter using a calibrated digital micrometer and recording the precise dimension (e.g., $25.042\text{ mm}$). This allows the team to calculate $\bar{x}$, $s$, and $C_{pk}$, pinpointing exactly how close the process operates to specification limits and detecting drift long before non-conforming scrap is produced.