Sampling and the central limit theorem
Populations vs samples, random sampling, standard error, and why averages behave normally.
We almost never measure every part. We measure a sample and use it to estimate the population (all parts the process makes, or will make).
Good samples
- Random: every unit has the same chance of being chosen.
- Representative: covers all shifts, machines, cavities and lots that matter.
- Large enough: enough units to estimate what you need with useful precision.
- In SPC, samples are taken in rational subgroups over time rather than at random, on purpose (see that article).
Standard error
The standard deviation of a sample mean. It shrinks as the sample grows, but only with the square root of n: four times the data halves the uncertainty.
The central limit theorem
For most distributions, the mean of n independent values is close to normally distributed when n is moderately large, with mean and standard deviation . Subgroup averages of 4 or 5 already look fairly normal unless the underlying data are very skewed. This is why Xbar charts are robust to non-normal data.
Sampling bias to avoid
- Measuring only the first parts after setup.
- Letting the operator choose "good-looking" parts.
- Sampling only on day shift.
- Dropping parts that were reworked.
Questions and answers
How many parts should I measure?
It depends on the decision. For a control chart, collect at least 20 to 25 subgroups before setting limits. For capability, at least 100 values over representative conditions. For comparing two processes, use a sample size calculation.
Is a bigger sample always better?
Bigger samples give more precise estimates, but measuring 1,000 parts from one hour tells you little about next week. Spread samples over time and conditions.
