9.1 Data Types, Measurement Scales & Sampling Techniques

9.1 Data Types, Measurement Scales & Sampling Techniques

Data is the lifeblood of continuous improvement and quality assurance. Without objective, verifiable data, organizations rely on guesswork, subjective opinion, and cognitive bias. In the words of W. Edwards Deming, "In God we trust; all others must bring data." On the ASQ Certified Quality Improvement Associate (CQIA) examination, quality professionals must demonstrate comprehensive fluency in the taxonomy of data, the four fundamental scales of measurement, probability sampling designs, sample size calculations, and the mechanics of acceptance sampling standards.

---

1. Data Classification: Qualitative / Attribute vs. Quantitative / Variable Data

In quality engineering and operational analytics, all collected information is divided into two primary categories: Attribute (Qualitative / Discrete) Data and Variable (Quantitative / Continuous) Data.

+-------------------------------------------------------------------------+
|                        QUALITY DATA TAXONOMY                            |
+-------------------------------------------------------------------------+
|                                                                         |
|                              [ ALL DATA ]                               |
|                                    │                                    |
|                  ┌─────────────────┴─────────────────┐                  |
|                  ▼                                   ▼                  |
|     [ ATTRIBUTE / QUALITATIVE ]             [ VARIABLE / QUANTITATIVE ] |
|     * Discrete counts / categories          * Continuous measurements   |
|     * Binary (Conforming / Defective)       * Infinite gradations       |
|     * Defect occurrences (Poisson)          * Measurable dimensions     |
|     * Counts of items (Binomial)            * Normal distribution       |
|                  │                                   │                  |
|         ┌────────┴────────┐                 ┌────────┴────────┐         |
|         ▼                 ▼                 ▼                 ▼         |
|    [ NOMINAL ]       [ ORDINAL ]       [ INTERVAL ]       [ RATIO ]     |
|    Names/Labels      Ranked order      Arbitrary 0        Absolute 0    |
|    (Pass/Fail)       (Low/Med/High)    (Temp °C / °F)     (mm, kg, sec) |
|                                                                         |
+-------------------------------------------------------------------------+

Qualitative / Attribute Data (Discrete Data)

• Definition: Data that is counted in whole, indivisible units or classified into distinct, mutually exclusive qualitative categories. Attribute data cannot be meaningfully divided into fractional or decimal subdivisions.
• Subtypes:
  • Binary / Dichotomous: Two possible states (e.g., Conforming vs. Nonconforming, Go vs. No-Go, Pass vs. Fail, Yes vs. No).
  • Nominal Categories: Unordered classes (e.g., Defect type: Scratch, Dent, Burred edge, Discoloration).
  • Count / Discrete Poisson Data: Whole counts of occurrences per unit area, volume, or time (e.g., 3 solder bridges per circuit board, 2 customer complaints per day).
• Underlying Probability Distributions: Binomial distribution (conforming/nonconforming proportions) and Poisson distribution (count of defects per unit).
• Statistical Control Charts: $p$-charts (proportion nonconforming), $np$-charts (number nonconforming), $c$-charts (count of defects for constant sample size), and $u$-charts (defects per unit for variable sample size).

Quantitative / Variable Data (Continuous Data)

• Definition: Data obtained by measuring physical, chemical, electrical, or temporal characteristics along an unbroken, continuous numerical scale. Variable data can take any real numerical value within a given measurement interval, constrained only by the resolution of the measuring instrument.
• Examples: Shaft outer diameter ($25.408\text{ mm}$), paint thickness ($12.5,\mu\text{m}$), process cycle time ($142.6\text{ seconds}$), tensile break force ($850.3\text{ N}$), operating temperature ($185.4,^\circ\text{C}$).
• Underlying Probability Distribution: Continuous Normal (Gaussian) distribution, log-normal, Weibull, or exponential distributions.
• Statistical Control Charts: $\bar{X}-R$ charts (Subgroup Average and Range), $\bar{X}-s$ charts (Subgroup Average and Standard Deviation), and $I\text{-}MR$ / $X\text{-}mR$ charts (Individual values and Moving Range).

Operational Comparison: Attribute vs. Variable Data

Key Principle: Variable data provides significantly higher statistical power and diagnostic insight than attribute data. Where feasible, continuous measurement is always preferred over binary classification because it enables proactive process control before nonconforming products are manufactured.

---

2. Stevens' Four Scales of Measurement (NOIR)

In 1946, psychologist and measurement theorist Stanley Smith Stevens published a landmark taxonomy classifying all data into four hierarchical measurement scales: Nominal, Ordinal, Interval, and Ratio (frequently remembered by the acronym NOIR). Each higher scale possesses all the properties of the preceding scale plus additional mathematical characteristics.

+-------------------------------------------------------------------------+
|                    STEVENS' MEASUREMENT HIERARCHY                       |
+-------------------------------------------------------------------------+
|                                                                         |
|  [ RATIO ]    ──► True Absolute Zero | All Math & Ratios Valid (mm, kg) |
|       ▲                                                                 |
|       │ (Adds meaningful ratios & true origin zero)                     |
|  [ INTERVAL ] ──► Equal Intervals | Arbitrary Zero (+ / - only) (°C, °F)|
|       ▲                                                                 |
|       │ (Adds equal, quantified distances between points)               |
|  [ ORDINAL ]  ──► Ranked Order | Unequal/Unknown Intervals (1st, 2nd)   |
|       ▲                                                                 |
|       │ (Adds logical progression and hierarchy)                        |
|  [ NOMINAL ]  ──► Identity / Labeling Only | No Order (Line A, Line B)  |
|                                                                         |
+-------------------------------------------------------------------------+

Detailed Analysis of the Four Scales

1. Nominal Scale (Naming / Classification)

• Mathematical Property: Identity and equivalence ($=$ or $\ne$). Numbers or words serve strictly as qualitative labels.
• Order: None. Swapping the order of categories has no mathematical meaning.
• Permissible Operations: Counting frequencies and percentages.
• Permissible Central Tendency: Mode only (Median and Mean are mathematically undefined and invalid).
• Quality Examples: Machine identifiers (Machine #1, Machine #2), Operator IDs, Defect types (Scratch, Crack, Stain), Plant locations (Dallas, Chicago, Atlanta).

2. Ordinal Scale (Order / Ranking)

• Mathematical Property: Identity and rank order ($=, \ne, >, <$). Elements can be arranged in a meaningful ascending or descending hierarchy.
• Interval Spacing: Unequal or undefined. The distance between 1st place and 2nd place is not necessarily equal to the distance between 2nd place and 3rd place.
• Permissible Central Tendency: Median and Mode (Arithmetic mean is technically invalid because distances between points are not quantified).
• Quality Examples: Customer satisfaction ratings (1 = Very Dissatisfied, 2 = Dissatisfied, 3 = Neutral, 4 = Satisfied, 5 = Very Satisfied); Supplier quality tiers (Bronze, Silver, Gold, Platinum); Mohs mineral hardness scale; Defect severity classifications (Minor, Major, Critical).

3. Interval Scale (Equal Intervals / Arbitrary Zero)

• Mathematical Property: Identity, rank order, and known, equal quantitative distances between adjacent points ($=, \ne, >, <, +, -$).
• Zero Point: Arbitrary (Relative). Zero does not represent the complete absence of the property being measured.
• Limitation: Ratios cannot be calculated. For example, $40,^\circ\text{C}$ is not "twice as hot" as $20,^\circ\text{C}$ because $0,^\circ\text{C}$ is merely the freezing point of water, not absolute thermal zero.
• Permissible Central Tendency & Dispersion: Mean, Median, Mode, Range, Variance, and Standard Deviation.
• Quality Examples: Temperature measured in Celsius or Fahrenheit; Calendar dates and years (e.g., the year 2026); Standardized test scores (e.g., SAT, IQ scores).

4. Ratio Scale (Absolute True Zero / Meaningful Ratios)

• Mathematical Property: Identity, rank order, equal intervals, and a true, non-arbitrary absolute zero point ($=, \ne, >, <, +, -, \times, \div$). Absolute zero represents the complete absence of the measured physical attribute.
• Ratios: Fully valid. A part weighing $10\text{ kg}$ is exactly twice as heavy as a part weighing $5\text{ kg}$, and $0\text{ kg}$ indicates zero mass.
• Permissible Central Tendency & Statistics: All statistical metrics, including Arithmetic Mean, Geometric Mean, Harmonic Mean, Standard Deviation, Coefficient of Variation, and advanced parametric statistical tests.
• Quality Examples: Length, width, thickness, mass, electrical current (Amperes), electrical resistance (Ohms), cycle time (seconds), force (Newtons), cost (dollars), and absolute thermodynamic temperature in Kelvin ($0\text{ K} = \text{absolute zero}$).

Comprehensive Measurement Scale Reference

---

3. Probability vs. Non-Probability Sampling Methods

Sampling is the process of selecting a representative subset of units from a larger target population to draw valid statistical conclusions about the entire population. In quality management, testing 100% of production is often economically prohibitive, logistically impossible, or physically destructive (e.g., tensile break testing, crash testing).

+-------------------------------------------------------------------------+
|                    COMMON PROBABILITY SAMPLING DESIGNS                  |
+-------------------------------------------------------------------------+
|                                                                         |
|  1. SIMPLE RANDOM SAMPLING                                              |
|     [ x ] [ . ] [ . ] [ x ] [ . ] ──► Every item has an equal and       |
|     [ . ] [ x ] [ . ] [ . ] [ x ]     independent chance (1/N) of draw. |
|                                                                         |
|  2. STRATIFIED SAMPLING                                                 |
|     Shift 1: [ x ] [ x ]          ──► Divide into homogeneous strata;   |
|     Shift 2: [ x ] [ x ]              sample proportionally from each   |
|     Shift 3: [ x ] [ x ]              subgroup to reduce variance.      |
|                                                                         |
|  3. SYSTEMATIC SAMPLING                                                 |
|     [ x ] . . [ x ] . . [ x ] . . ──► Select every k-th item from a     |
|                                       sequential stream (k = N/n).      |
|                                                                         |
|  4. CLUSTER SAMPLING                                                    |
|     Box 1: [ . . . ]  Box 2: [ x x x ] ──► Randomly pick entire clusters|
|     Box 3: [ . . . ]  Box 4: [ x x x ]     and inspect all units inside.|
|                                                                         |
+-------------------------------------------------------------------------+

Probability Sampling Methods

In probability sampling, every element in the population has a known, non-zero probability of being selected. This provides an unbiased foundation for statistical inference.

1. Simple Random Sampling (SRS):
   • Mechanism: Every individual unit in the population ($N$) has an equal and independent probability ($1/N$) of selection.
   • Implementation: Generating random numbers via computer algorithms or random number tables matched against serialized production logs.
   • Advantage: Unbiased; mathematically straightforward.
   • Disadvantage: Requires a complete, numbered sampling frame of the entire population; can be costly and logistically inefficient over wide geographic areas.
2. Stratified Random Sampling:
   • Mechanism: The population is divided into distinct, mutually exclusive, homogeneous subgroups called strata (e.g., production shifts, manufacturing lines, raw material vendor lots). A simple random sample is then drawn independently from each stratum, usually proportional to the stratum's size relative to the population.
   • Advantage: Guarantees representation of all subgroups, reduces within-stratum variance, and yields more precise population estimates than simple random sampling.
   • Quality Application: Ensuring sample lots include parts produced by Day, Swing, and Graveyard shifts to account for shift-to-shift variation.
3. Systematic Sampling:
   • Mechanism: Selecting every $k\text{-th}$ item from a sequential process or list after a random starting point between $1$ and $k$, where the sampling interval is $k = N / n$.
   • Advantage: Extremely easy to execute on active manufacturing lines (e.g., inspecting every 25th bottle coming off a filling line).
   • Critical Hazard (Periodicity / Cyclic Bias): If the sampling interval $k$ coincides with a cyclic physical phenomenon in the machinery (e.g., an 8-cavity injection mold where every 8th part comes from the same defective cavity), the sample will be severely biased.
4. Cluster Sampling:
   • Mechanism: The population is divided into naturally occurring heterogeneous groups or clusters (e.g., shipping pallets, shipping containers, physical bins). A random sample of clusters is selected, and all units (or a random subsample) within the selected clusters are inspected.
   • Advantage: Highly cost-effective and logistically practical for bulk shipments or geographically dispersed sites.
   • Disadvantage: Higher sampling error (lower statistical efficiency) compared to SRS because units within the same cluster often share similar localized characteristics.

Non-Probability Sampling (Subject to Severe Bias)

• Convenience Sampling: Selecting units that are easiest to access (e.g., grabbing only the top 10 parts from a deep pallet bin). Introduces severe physical and temporal bias.
• Judgment / Purposive Sampling: An auditor or inspector selects parts based on subjective intuition or personal experience. Cannot be generalized statistically.
• Quota Sampling: Non-random selection of units to fulfill predetermined category quotas.

Exam Warning: Non-probability sampling methods do not satisfy statistical assumptions of randomness. They cannot be used to calculate valid confidence intervals, process capability indices, or margin of error.

---

4. Sampling Bias, Sampling Error & Sample Size Considerations

Types of Sampling and Non-Sampling Errors

• Sampling Error: The natural statistical difference between a sample statistic (e.g., sample mean $\bar{x}$) and the true population parameter (population mean $\mu$) resulting from observing only a portion of the population. Sampling error is inevitable but decreases predictably as sample size increases.
• Non-Sampling Error (Systemic Bias): Systematic distortion caused by flawed sampling design, defective measurement instruments, poor operational definitions, or data transcription errors. Unlike sampling error, non-sampling error cannot be reduced merely by increasing sample size.
• Common Biases:
  • Selection Bias: Systematic exclusion or overrepresentation of specific subsets of the population.
  • Survivorship Bias: Analyzing only items that survived a screening or stress-testing stage, ignoring failed units.
  • Non-Response Bias: Systematic differences between individuals who respond to surveys and those who refuse or ignore them.

Sample Size Determination Mechanics

To calculate the required sample size ($n$) for estimating a population mean with continuous (variable) data, quality analysts must specify three key statistical parameters:

1. Confidence Level ($1 - \alpha$): The desired probability that the true population mean falls within the calculated interval, corresponding to the critical standard normal value $Z_{\alpha/2}$ (e.g., $Z = 1.96$ for $95%$ confidence, $Z = 2.576$ for $99%$ confidence).
2. Margin of Error ($E$): The maximum permissible difference between the sample mean and the true population mean ($E = |\bar{x} - \mu|$).
3. Population Variability ($\sigma$): The standard deviation of the population (or an estimate $s$ from historical pilot data).

$\text{Sample Size Formula (Continuous Data):} \quad n = \left( \frac{Z_{\alpha/2} \cdot \sigma}{E} \right)^2 = \frac{Z_{\alpha/2}^2 \cdot \sigma^2}{E^2}$

The Square Root Law of Sampling

The standard error of the mean (the standard deviation of sample means) is defined as:

$\sigma_{\bar{x}} = \frac{\sigma}{\sqrt{n}}$

+-------------------------------------------------------------------------+
|                    THE SQUARE ROOT LAW OF SAMPLING                      |
+-------------------------------------------------------------------------+
|                                                                         |
|   To reduce the Standard Error (Margin of Error) by:                    |
|                                                                         |
|   * 50% (Cut in half, 1/2)    ──► Must QUADRUPLE sample size (4x n)     |
|   * 66.7% (Cut to 1/3)        ──► Must INCREASE sample size by 9x n     |
|   * 75% (Cut to 1/4)          ──► Must INCREASE sample size by 16x n    |
|   * 90% (Cut to 1/10)         ──► Must INCREASE sample size by 100x n   |
|                                                                         |
+-------------------------------------------------------------------------+

---

5. Acceptance Sampling & Statistical Quality Standards

Acceptance Sampling is a quality inspection methodology used to make a lot sentencing decision (accepting or rejecting an entire production lot) based on the inspection of a statistically determined sample of units from that lot. Acceptance sampling is an audit and sentencing tool—it does not directly control process variation or improve product quality; it merely acts as an economic filter.

+-------------------------------------------------------------------------+
|                 OPERATING CHARACTERISTIC (OC) CURVE                     |
+-------------------------------------------------------------------------+
|  Probability of                                                         |
|  Acceptance Pa                                                          |
|   1.0 ┤  (1 - α) ──► Producer's Risk (Type I Error: Reject Good Lot)    |
|       │ ╲                                                               |
|       │   ╲   [AQL] (Acceptable Quality Limit)                          |
|       │     ╲                                                           |
|       │       ╲                                                         |
|       │         ╲                                                       |
|       │           ╲                                                     |
|   0.1 ┤             ╲   [LTPD / RQL] (Lot Tolerance Percent Defective)  |
|       │               ╲  β ──► Consumer's Risk (Type II: Accept Bad)    |
|   0.0 └─────────────────┴──────────────────────────────────────────────► |
|       0%               AQL               LTPD                     % Def |
+-------------------------------------------------------------------------+

Core Acceptance Sampling Terminology

• Acceptable Quality Limit (AQL): The poorest level of quality for the supplier's process that the consumer considers satisfactory as a process average. Associated with Producer's Risk ($\alpha$, Type I Error)—the risk of erroneously rejecting an acceptable lot (typically set at $\alpha = 0.05$ or $5%$).
• Lot Tolerance Percent Defective (LTPD / Rejectable Quality Level [RQL]): The designated high defect level that the consumer wishes to reject consistently. Associated with Consumer's Risk ($\beta$, Type II Error)—the risk of erroneously accepting a nonconforming lot (typically set at $\beta = 0.10$ or $10%$).
• Average Outgoing Quality (AOQ): The expected average quality level of outgoing product over a long sequence of lots, assuming rejected lots undergo $100%$ rectifying inspection (sorting and replacement of defective parts).
• Average Outgoing Quality Limit (AOQL): The absolute maximum percentage of defective items that can exist in the outgoing product over the long run, regardless of incoming quality.

Major Acceptance Sampling Systems

1. ANSI/ASQ Z1.4 (ISO 2859-1 / Former MIL-STD-105E)

• Focus: Attribute Sampling indexed primarily by AQL.
• Structure: Single, double, and multiple sampling plans.
• Inspection Levels: General Inspection Levels (I, II, III, with Level II as standard) and Special Inspection Levels (S-1, S-2, S-3, S-4 for destructive/costly tests).
• Dynamic Switching Rules:
  • Normal Inspection: Default starting mode.
  • Tightened Inspection: Triggered when $2$ out of $5$ consecutive lots are rejected under normal inspection.
  • Reduced Inspection: Permitted when $10$ consecutive lots are accepted under normal inspection and production is stable.
  • Discontinuation: If $5$ consecutive lots remain on tightened inspection, sampling is discontinued until supplier corrective actions are verified.

2. ANSI/ASQ Z1.9 (ISO 3951 / Former MIL-STD-414)

• Focus: Variable Sampling indexed by AQL. Uses sample mean and standard deviation to calculate distance to specification limits. Requires much smaller sample sizes than Z1.4 for equivalent statistical protection.

3. Dodge-Romig Sampling Plans

• Focus: Developed by Harold Dodge and Harry Romig for internal manufacturing inspection.
• Orientation: Indexed by LTPD (Consumer protection) and AOQL (Average Outgoing Quality Limit).
• Rectifying Inspection: Assumes that every rejected lot undergoes $100%$ screening inspection where all defective items are removed and replaced with conforming units.