Mean vs. Median: Which Should You Choose to Manage Your Process?
When analyzing data from a manufacturing process, the choice of a measure of central tendency—mean or median—may seem trivial. In reality, it determines the reliability of your conclusions, the relevance of your control charts, and the accuracy of your capability indices. Understanding when to use one rather than the other is a fundamental skill for any Lean Six Sigma practitioner.
What These Two Indicators Actually Measure
Visit average (or arithmetic mean) is the sum of all the values divided by their number. It incorporates information from each data point, making it a very effective estimator when the data follow a normal distribution—that is, a symmetric distribution without outliers.
Visit median is the value located in the middle of a dataset sorted in ascending order. It is called sturdy : An outlier—even one that is far removed from the rest of the observations—shifts the distribution only slightly. On the other hand, it uses less information than the mean and may be less accurate for small samples.
In practice, the two coincide perfectly in a perfectly symmetric distribution. It is when they diverge that the information becomes useful: their difference indicates an asymmetry or the presence of outliers in the process.
When the Average Is the Right Choice
The average is the benchmark indicator as long as your data meets two conditions: normality (or a sufficiently symmetric distribution) and absence of unjustified outliers. That is why it is central to the most widely used statistical tools in statistical process control (SPC):
- Visit X-bar/R control charts and X-bar / S are based on the average of the subgroups to detect deviations.
- Visit Capability indices Cp and Cpk use the mean as an estimator of the process position.
- Visit t-tests and the ANOVA (analysis of variance) relies on means to compare populations.
If your process is under control and produces normally distributed data, the mean takes into account all the available information and provides the most accurate estimators in the statistical sense.
When the Median Takes Center Stage
The median becomes preferable—or essential—in three common situations in an industrial setting.
Asymmetric or long-tail data
Certain processes naturally generate asymmetric distributions: cycle times, contamination levels, and delivery times. A single, exceptionally long operation can skew the average upward and provide a distorted picture of the process’s typical behavior. The median, therefore, better reflects the actual experience of the operator or customer.
Outliers identified
When a sensor drifts temporarily or an exceptional event (machine shutdown, non-conforming raw material) introduces outliers into the data set, the median mitigates their effect. This does not eliminate the need to investigate and correct the root cause, but it prevents the process from being controlled based on an artificially distorted statistic.
Small samples in ordinal data
For attributes rated on an ordinal scale (satisfaction scores, severity levels of a defect), the mean can be calculated mathematically but is semantically questionable. The median retains a concrete meaning: it denotes the actual midpoint of the sample.
There are also median control charts (X̃/R charts), which are less sensitive than X-bar charts but more robust when dealing with non-normal data or operators who are unfamiliar with calculating means.
How to determine which one to use with your data
Best practice is not to choose at random, but to analyze the distribution before making a decision.
- Plot a histogram or a box plot (box plot) : Visible asymmetry, a long tail, or isolated spots are warning signs.
- Compare the mean and the median : A deviation of more than 10–15 % from the data range warrants attention.
- Perform a normality test (Shapiro-Wilk for small samples, Anderson-Darling for large ones): if the normality assumption is rejected, the median is generally more representative.
- Check for outliers using a Grubbs test or an individual control chart: if they can be attributed to a specific cause, isolate them before calculating your indicator.
These steps take just a few minutes in any statistical analysis software and help prevent management decisions based on inappropriate metrics.
The question is not whether the mean is «better» than the median: each is appropriate for a specific context. For a stable, normally distributed process, the mean remains the most powerful estimator. As soon as asymmetries or outliers come into play, the median provides a more accurate picture of the reality on the ground. Diagnose first, then choose: this is the rigorous approach that distinguishes the Six Sigma practitioner from those who apply formulas without questioning their assumptions.