6.4 Sampling Concepts, Methods, and Sample Size for Estimation
Key Takeaways
- A formal Data Collection Plan specifies operational definitions, measurement instruments, data types, sampling methodology, and sample frequency to eliminate collection bias.
- Stratified Sampling partitions a population into homogeneous strata (e.g., shifts, lines, suppliers) and samples randomly within each, reducing overall sampling variance.
- Systematic Sampling selects every kth item from a continuous process flow; practitioners must audit for cyclical process periodicities that match the sampling interval.
- Sample size formulas establish the minimum sample required to estimate a mean (n = (z * s / E)^2) or proportion (n = (z^2 * p * (1 - p)) / E^2) for a given confidence level and margin of error.
Data collection is the critical bridge in the Measure phase connecting defined operational problems to empirical statistical analysis. Flawed data collection invalidates subsequent statistical inferences, regardless of how advanced the analytical tools are. A Black Belt must design a structured Data Collection Plan and execute rigorous probability sampling strategies to capture representative, unbiased process data.
Probability Sampling Methodologies
Sampling is the process of selecting a subset ($n$) from a population universe ($N$) to make statistical inferences about population parameters. Non-probability sampling (such as convenience or judgment sampling) introduces severe selection bias and renders inferential statistics invalid. Six Sigma requires probability sampling:
| Sampling Method | Operational Procedure | Primary Advantage | Major Risk / Pitfall |
|---|---|---|---|
| Simple Random Sampling | Every item in the population frame has an equal, independent probability of selection (e.g., random number generator). | Completely unbiased; straightforward statistical math. | Requires a complete, accessible sampling frame of the entire population. |
| Stratified Sampling | Population is divided into homogeneous non-overlapping sub-groups (strata) based on key factors (e.g., shifts, production lines, raw material lots). Random samples are drawn within each stratum. | Ensures representation across sub-groups; reduces sampling variance compared to simple random sampling. | Requires prior knowledge of relevant stratification factors. |
| Systematic Sampling | Selects every $k^{\text{th}}$ item from a sequential process stream, where sampling interval $k = \frac{N}{n}$. | Easy to execute in continuous manufacturing or service flows. | If hidden cyclical periodicities or patterns match interval $k$, samples will be severely biased. |
| Cluster Sampling | Population is divided into heterogeneous clusters (e.g., shipping pallets, branch offices). Random clusters are selected, and all items within chosen clusters are evaluated. | Cost-effective for geographically dispersed or bulk-packaged items. | Higher sampling error if clusters are internally homogeneous rather than heterogeneous. |
Sample Size Determination Formulas
Selecting an arbitrary sample size (such as "sample 30 parts") without statistical justification leads to either under-powered studies (failing to detect real effects) or wasteful over-sampling. Sample size equations depend on whether the metric is continuous or discrete.
1. Sample Size for Estimating a Continuous Population Mean ($\mu$)
To estimate a population mean within a specified allowable margin of error $E$ at a confidence level of $(1 - \alpha)$:
Where:
- $n$ = Required minimum sample size (always rounded UP to the nearest integer).
- $z_{\alpha/2}$ = Critical value from the standard normal distribution corresponding to the desired confidence level ($z_{0.025} = 1.96$ for $95%$ confidence; $z_{0.005} = 2.576$ for $99%$ confidence).
- $\sigma$ = Estimated population standard deviation (derived from historical baseline data or pilot study sample standard deviation $s$).
- $E$ = Maximum allowable margin of error (precision) in the original units of measurement.
Example Calculation: A Black Belt wants to estimate mean shaft diameter within a margin of error $E = 0.02\text{ mm}$ at a $95%$ confidence level ($z = 1.96$). Historical data shows process standard deviation $\sigma = 0.10\text{ mm}$. Rounding up, the team must sample at least 97 parts.
2. Sample Size for Estimating a Discrete Population Proportion ($p$)
To estimate a population defect rate or proportion within a margin of error $E$ at a confidence level of $(1 - \alpha)$:
Where:
- $p$ = Estimated baseline population proportion. If unknown, use $p = 0.50$, which maximizes the product $p(1 - p) = 0.25$ and yields the most conservative (largest) sample size guarantee.
- $E$ = Maximum allowable margin of error expressed as a decimal proportion (e.g., $E = 0.03$ for $\pm 3%$).
Example Calculation: Estimate the invoice error proportion within a margin of error $E = 0.04$ ($4%$) at $95%$ confidence ($z = 1.96$), with no prior proportion knowledge ($p = 0.50$). Rounding up, the team must sample at least 601 invoices.
Sampling Concepts: Representativeness, Homogeneity, Bias, Accuracy, and Precision
The Body of Knowledge asks Black Belts to define and describe sampling concepts and then to determine the appropriate sampling method for a given situation, at the Evaluate level. Five concepts underpin that judgment.
Representative selection. A sample is representative when every relevant characteristic of the population appears in it in roughly the population's proportions. The practical test is not how the sample was drawn in the abstract but whether the known sources of variation -- shift, machine, lot, product family, time of day -- are all present in it. A sample of 200 units drawn entirely from Tuesday day shift is large and unrepresentative.
Homogeneity. A population is homogeneous when its members come from the same underlying process. When they do not -- two machines, two suppliers, two shifts with different methods -- the population is a mixture, and a single sample statistic describes an average that no individual member resembles. Mixtures show up as bimodal histograms, points avoiding the centre line on a control chart, and capability indices that are stable yet inexplicably poor. The remedy is to stratify and sample each stratum separately.
Bias. A systematic error that pushes the estimate away from the truth in a consistent direction. It is not reduced by taking a larger sample, which is what makes it more dangerous than random error.
| Bias type | Mechanism | Example |
|---|---|---|
| Selection bias | Some population members cannot be selected | Sampling only units that reached final inspection |
| Convenience bias | Sampling what is easy to reach | Taking parts from the top of the bin |
| Non-response bias | Those who do not respond differ from those who do | Survey answered mainly by dissatisfied customers |
| Measurement bias | The instrument or method reads consistently high or low | Uncalibrated gauge; leading survey question |
| Time bias | The sampling window is unrepresentative | Baseline taken only during a seasonal peak |
Accuracy is closeness to the true value -- the absence of bias. Precision is closeness of repeated measurements to one another -- small random variation. They are independent: a sample can be precise and badly biased, which is the worst combination because the consistency makes the result look trustworthy. Accuracy is addressed by calibration and by sampling design; precision is addressed by sample size and by gauge repeatability.
Choosing a sampling method
| Method | How it works | Best when | Weakness |
|---|---|---|---|
| Simple random | Every member has an equal probability of selection | The population is homogeneous and a complete frame exists | Requires a frame; may miss small strata by chance |
| Stratified | Divide into homogeneous strata, then sample within each | Known subgroups differ -- shift, machine, product family | Requires knowing the strata in advance |
| Systematic | Take every $k$th unit, with a random start | Continuous production; simple to administer | Fails badly if the sampling interval coincides with a process cycle |
| Subgroup (rational) | Small samples of consecutive units at intervals | Control charting; separates within-subgroup from between-subgroup variation | Deliberately not representative of long-term variation |
| Block / cluster | Sample whole natural groups such as a pallet, batch, or shift | The population is naturally clustered and sampling individuals is costly | Higher variance than simple random for the same $n$ |
Two selection rules carry most of the practical weight. Stratify whenever you know the population is not homogeneous -- it both protects the estimate and reveals the between-stratum difference, which is usually the most valuable finding. And avoid systematic sampling when the process has any cyclical pattern: taking every 12th unit from a 12-cavity mould samples one cavity forever, and taking a reading every 60 minutes on an hourly maintenance cycle reproduces the same phase every time.
Rational subgrouping deserves a separate note because it is deliberately the opposite of representative: the subgroup is chosen so that only common-cause variation appears within it, and all other variation appears between subgroups, where the control chart can detect it.
A Six Sigma team needs to estimate the average filled weight of cereal boxes produced on an automated packaging line. Historical pilot studies indicate a standard deviation of 4.0 grams. The team wants to estimate the true mean within an allowable margin of error of 0.8 grams at a 95% confidence level (z = 1.96). What is the minimum required sample size?
A manufacturing plant operates three distinct work shifts (Day, Evening, Night). Historical quality reports show that variation differs significantly between shifts due to ambient room temperature changes. Which sampling method ensures that sample data accurately reflects all shifts while minimizing overall sampling variance?
What primary risk must a Black Belt evaluate before implementing systematic sampling (sampling every kth part) on a continuous automated assembly line?