4.1 Data Types, Sampling Methods, and Populations

4.1 Data Types, Sampling Methods, and Populations

Introduction to Quality Metrology and Statistical Inference

In modern manufacturing and technical inspection, decisions regarding product conformance, machine setup, and process stability cannot be made on subjective impressions. Quality technicians must collect, structure, and interpret physical measurements and observational counts. Statistical inference is the scientific discipline of drawing valid conclusions about a large body of items (the population) by inspecting only a smaller, representative subset (the sample).

Before calculating control limits, evaluating process capability, or deciding whether to accept or quarantine a production lot, a certified technician must understand what type of data is being gathered. The underlying mathematical scale dictates which statistical formulas are valid, how large the inspection sample must be, and what conclusions can be defended during an internal or regulatory audit.

---

Qualitative vs. Quantitative Data

At the most fundamental taxonomic level, quality data divides into two primary categories:

1. Qualitative Data (Categorical Data): Non-numerical descriptions, labels, or subjective classifications. These describe qualities or attributes of an item rather than measuring an amount. Examples include alloy type (316L stainless steel vs. 6061-T6 aluminum), defect classification (burr, crack, scratch, contamination), surface coating color (black oxide, clear anodize), or supplier designation.
2. Quantitative Data (Numerical Data): Objective numerical values representing counts or physical measurements. Arithmetic operations such as addition, subtraction, averaging, and standard deviation calculations are meaningful for quantitative data. Examples include tensile yield strength (450 MPa), shaft diameter (25.402 mm), hole depth (12.5 mm), and the number of voids per square centimeter.

---

Discrete (Attribute) vs. Continuous (Variable) Data

Within quantitative and operational data, the ASQ Body of Knowledge draws a crucial distinction between discrete (attribute) data and continuous (variable) data. This distinction directly determines whether a technician uses $p$, $np$, $c$, or $u$ control charts (for attributes) or $\bar{X}-R$ and $\bar{X}-s$ charts (for variables).

Discrete (Attribute) Data

Discrete data consists of distinct, separate values that can only take on specific integer numbers or binary categories. Discrete data is counted; it cannot be infinitely subdivided into fractions.

• Binary (Go / No-Go) Data: The part either conforms or does not conform. A precision plug gage either enters a reamed bore (Go) or jams (No-Go). A circuit breaker either trips under test current (Pass) or fails to open (Fail).
• Count Data (Poisson Type): The number of distinct defects observed per unit or inspection area. For example: 3 solder bridges on a printed circuit board, 0 scratches on a painted hood, or 4 gas pores in an aerospace weld radiographic image. You cannot have 2.37 distinct solder bridges on a single board.

Continuous (Variable) Data

Continuous data represents physical measurements along an unbroken numerical scale. Between any two recorded continuous values, an infinite number of intermediate values theoretically exists. Continuous data is limited only by the resolution and precision of the measuring instrument (micrometer, height gage, optical comparator, or coordinate measuring machine).

• Examples include outer diameter measured with a digital micrometer (1.5032 in), Rockwell C hardness (58.4 HRC), curing oven temperature (182.6 °C), plating thickness (12.4 μm), and mass (145.22 g).

Comparison: Attribute vs. Variable Data in Inspection

[!IMPORTANT] ASQ Core Rule: Variable data is always preferred over attribute data whenever economically feasible. Because continuous data reveals where parts fall within specification limits, it provides early warning of tool wear and machine thermal drift long before a Go/No-Go gage rejects a scrap part.

---

The Four Levels of Measurement (Stevens' Scales)

In 1946, psychologist S. S. Stevens established the classic taxonomy classifying data into four hierarchical levels of measurement: Nominal, Ordinal, Interval, and Ratio. Moving from Nominal to Ratio, each successive level retains all properties of the lower levels while adding new mathematical properties.

+---------------------------------------------------------------------------------+
|                       STEVENS' HIERARCHY OF MEASUREMENT                         |
+----------+--------------------+--------------------------------+----------------+
| Scale    | Core Property      | Permissible Math / Statistics  | Quality Example|
+----------+--------------------+--------------------------------+----------------+
| Ratio    | Absolute True Zero | +, -, *, /, Mean, StDev, CV    | Length, Weight |
| Interval | Arbitrary Zero     | +, -, Mean, StDev (No ratios)  | Temp (°C, °F)  |
| Ordinal  | Ordered Ranking    | Greater/Less than, Median, IQR | Defect Severity|
| Nominal  | Unordered Labels   | Equality (=, ≠), Mode, Counts  | Machine #, Part|
+----------+--------------------+--------------------------------+----------------+

1. Nominal Scale

• Definition: Data consists of mutually exclusive categories or labels with no inherent natural order, mathematical ranking, or physical distance between them.
• Permissible Operations: Counting frequencies, percentages, mode. Arithmetic operations like addition or averaging are mathematically meaningless.
• Quality Applications: Tracking machine serial numbers (Lathe #1, Lathe #2), shift codes (Day, Swing, Graveyard), defect categories (Pinhole, Scratch, Denter, Burr), and material lots.

2. Ordinal Scale

• Definition: Data consists of categories that possess a natural, logical order or rank, but the intervals or distances between adjacent ranks are unequal, unknown, or subjective.
• Permissible Operations: Median, percentiles, rank correlation, greater-than / less-than comparisons. You cannot compute a valid arithmetic mean or standard deviation because a step from Rank 1 to Rank 2 does not equal a step from Rank 2 to Rank 3.
• Quality Applications: Defect severity ratings (Minor, Major, Critical), surface finish visual comparators (Rough, Moderate, Fine), supplier approval tiers (Gold, Silver, Bronze), and Likert-scale operator survey ratings (1 = Strongly Disagree to 5 = Strongly Agree).

3. Interval Scale

• Definition: Data possesses ordered ranks and constant, equal intervals between successive scale units, but has an arbitrary zero point. Zero does not signify the complete physical absence of the measured characteristic.
• Permissible Operations: Addition, subtraction, arithmetic mean, range, standard deviation. However, calculating multiplication ratios is invalid: 40 °C is not 'twice as hot' as 20 °C because 0 °C is an arbitrary freezing point of water, not the absolute cessation of molecular thermal kinetic energy.
• Quality Applications: Temperature in degrees Celsius (°C) or Fahrenheit (°F), calendar dates, and tooling offset datum shifts relative to an arbitrary machine zero.

4. Ratio Scale

• Definition: The highest and most mathematically powerful level of measurement. Ratio data possesses ordered ranks, constant equal intervals, and a true, non-arbitrary absolute zero point representing the complete absence of the physical quantity.
• Permissible Operations: All arithmetic operations (addition, subtraction, multiplication, division), geometric mean, standard deviation, and the Coefficient of Variation (CV). True physical ratios are valid: a pin measuring 10.00 mm is exactly twice as long as a 5.00 mm pin; a component weighing 0.000 g possesses zero mass.
• Quality Applications: Dimensional lengths, widths, diameters, wall thicknesses (mm, in); mass and weight (kg, lb); electrical resistance (ohms); tensile force (newtons); absolute temperature in Kelvin (K); cycle time (seconds); and leak rates (sccm).

---

Populations vs. Samples: Parameters vs. Statistics

A central duty of the quality technician is to avoid confusing the properties of an entire population with those of an inspected sample.

Population and Parameters

• Population: The complete collection of all items, parts, assemblies, or observations under consideration. It can be a finite lot (e.g., a heat-treat batch of $N = 2,500$ forged crankshafts) or an infinite stream (e.g., continuous stamping production running over three shifts).
• Parameter: A fixed numerical characteristic describing the entire population. Parameters are typically unknown and unobservable unless 100% of the population is measured with error-free gages. Population parameters are denoted using Greek letters.

Sample and Statistics

• Sample: A subset of units selected from the population according to a defined sampling plan. It is denoted by sample size $n$.
• Statistic: A numerical value computed from sample observations. Sample statistics serve as point estimates of the unknown population parameters. Sample statistics are denoted using Latin (English) letters.

Master Notation Table: Parameters vs. Statistics

---

Probability Sampling Methods in Quality Inspection

For sample statistics to provide unbiased, statistically defensible estimates of population parameters, the sample must be selected using a probability sampling method. In a probability sample, every unit in the population has a known, non-zero probability of selection.

1. Simple Random Sampling (SRS)

• Mechanism: Every individual unit in the population of size $N$ has an equal and independent probability of being chosen in the sample of size $n$.
• Implementation: Units are assigned sequential numbers from 1 to $N$. A random number generator or published random number table generates $n$ unique values, and the corresponding physical parts are pulled.
• Application: Static warehouse lots, receiving inspection of boxed fasteners, and final packaging lot audit.

2. Stratified Random Sampling

• Mechanism: The population is divided into mutually exclusive, homogeneous subgroups called strata based on an identifiable operational variable. A simple random sample is then drawn independently from each stratum, typically proportional to the stratum's size.
• Why Use It?: Stratified sampling significantly reduces sampling error compared to simple random sampling when there is high variability between strata but low variability within each stratum. It guarantees that every machine head, die cavity, or raw material heat lot is fairly represented in the sample.
• Application: Multi-cavity injection molding (sampling parts from each of the 16 distinct mold cavities), multi-spindle screw machines (sampling parts from each of the 6 spindles), or multi-shift production runs.

3. Systematic Sampling

• Mechanism: Units are selected at a constant, fixed periodic interval through time, space, or sequence. Every $k$-th unit is selected from the production stream, where the sampling interval is $k = N / n$. The initial starting point between 1 and $k$ must be chosen randomly.
• Application: High-speed packaging or bottle-filling lines where an inspector pulls every 50th container off the conveyor belt for tare-weight verification.
• The Critical Hazard — Periodicity: If the manufacturing process has an underlying cyclical rhythm or mechanical oscillation that matches the sampling interval $k$, systematic sampling introduces severe bias. For example, if a 4-station rotary dial indexing table has a worn bushing at Station 3, and the technician samples every 4th part ($k = 4$), the sample will either consist entirely of parts made on defective Station 3 or completely miss Station 3 forever!

4. Cluster Sampling

• Mechanism: The population is divided into naturally occurring heterogeneous groups called clusters (e.g., shipping crates, master cartons, or pallet totes). A random sample of clusters is selected, and either all units within the selected clusters are inspected (one-stage cluster sampling) or a random sample is drawn from the selected clusters (two-stage cluster sampling).
• Application: Receiving inspection where unpacking individual units across 100 master shipping crates is logistically prohibitive. The technician randomly selects 5 crates and inspects all units inside those 5 crates.
• Limitation: While cluster sampling minimizes material handling and logistics costs, it generally produces higher sampling error than simple random sampling if parts within a single crate are more similar to each other than to parts in other crates.

---

Non-Sampling Errors and Sampling Bias

When statistical conclusions fail to reflect production reality, the root cause traces to either sampling bias or non-sampling errors.

Sampling Bias (Selection Bias)

Sampling bias occurs when the physical method of collecting the sample systematically favors certain units over others, violating the principle of equal opportunity. The resulting sample statistic does not converge toward the true population parameter.

• Convenience Sampling: The most rampant quality inspection error on the shop floor. A technician pulls 10 parts from the top layer of a gaylord or pallet box because digging to the middle or bottom is difficult. If the tote filled sequentially, top parts represent only the final 10 minutes of a 4-hour production run, masking thermal warmup drift or tool wear occurring earlier.
• Judgment Sampling: An inspector relies on personal intuition to select 'typical' parts, intentionally discarding parts that look slightly rough. This artificially suppresses the observed sample standard deviation, hiding true process variability.

Non-Sampling Errors

Non-sampling errors are distortions that arise during data collection, inspection execution, or record recording, unrelated to the statistical sampling design. Non-sampling errors cannot be reduced by simply increasing sample size; they can only be eliminated through standardized procedures, metrology control, and training.

• Measurement System Error: Gage calibration drift, mechanical hysteresis, thermal expansion differences between steel gage blocks and aluminum workpieces, and poor operator gage repeatability and reproducibility (GR&R).
• Transcription and Data Entry Blunders: An inspector reads 12.045 mm on a digital bore gage but manually logs 12.450 mm on the paper traveler.
• Specification Ambiguity: Flawed blueprints, outdated revision drawings, or vague visual inspection criteria that cause different inspectors to classify identical parts inconsistently.

---

Real-World Shop-Floor Scenarios & Common Exam Traps

Shop Scenario: Multi-Spindle Screw Machine Sampling

A precision machining shop operates a 6-spindle automatic screw machine producing brass hydraulic fittings. The quality plan mandates inspecting 12 fittings every hour. The operator grabs the first 12 parts falling into the collection chute at the top of the hour.

• Metrology Diagnosis: This is an uncontrolled convenience sample. The 12 consecutive parts likely emerged from just two spindle cycles. Spindle #4, which may have a chipped grooving insert, might not be represented in the sample at all.
• Correct Quality Solution: Implement stratified random sampling. The technician must collect 2 parts directly from each of the 6 individual spindles every hour. This guarantees full representation of all spindle tooling geometries and separates spindle-to-spindle variation from within-spindle variation.

Common Exam Traps for CQT Candidates

• Exam Trap 1: Confusing Interval and Ratio Scales: The classic exam question asks whether temperature in degrees Celsius (°C) is an interval or ratio scale. Remember: Celsius and Fahrenheit are Interval scales because 0 °C is an arbitrary reference point (the freezing point of water). Kelvin is a Ratio scale because 0 K represents absolute zero. Length, weight, force, and resistance are Ratio data.
• Exam Trap 2: Believing Larger Samples Fix Convenience Bias: If a technician takes a biased sample (e.g., sampling only parts produced during the first 15 minutes of a shift), increasing the sample size from $n = 10$ to $n = 100$ from that same 15-minute window does not eliminate bias. It merely yields a highly precise estimate of a non-representative sub-population.
• Exam Trap 3: Confusing Greek and Latin Symbols: Examination questions frequently test parameter vs. statistic notation. Remember: $\mu$ and $\sigma$ are population parameters; $\bar{x}$ and $s$ are sample statistics. Stating that 'the sample mean is $\mu = 25.4$' is a fundamental notation error.