Chauvenet's Criterion — Definition, Formula & Example
Chauvenet's criterion is a rule for deciding whether one suspicious measurement in a data set is so far from the rest that you should discard it. You reject the point if fewer than half a measurement would be expected to land that far from the mean.
Let be measurements assumed to come from a normal distribution, with sample mean and sample standard deviation . For a suspect value , compute and the two-tailed probability , where is standard normal. The value is rejected if , or equivalently if .
Key Formula
Where:
- = Number of measurements, including the suspect one
- = The suspect measurement
- = Sample mean of all n measurements
- = Sample standard deviation of all n measurements
- = Distance of the suspect value from the mean, in standard deviations
- = Two-tailed standard normal probability of a deviation at least this large
Worked Example
Problem: You measure the length of a board 10 times and record these values in inches: 20.5, 19.7, 19.9, 19.9, 20.0, 20.0, 20.0, 20.0, 20.0, 20.0. Use Chauvenet's criterion to decide whether the reading 20.5 in. should be rejected.
Step 1: Find the mean and sample standard deviation of all 10 readings. Keep the suspect value in both. The deviations from the mean are 0.5, −0.3, −0.1, −0.1, and six zeros.
Step 2: Convert the suspect reading to a z-score. Then find the two-tailed probability of a deviation at least that large, using .
Step 3: Multiply by to get the expected number of readings this extreme, and compare it with 1/2. As a shortcut, for the cutoff is , which corresponds to .
Answer: Since , reject the 20.5 in. reading. The mean of the remaining nine readings is about 19.94 in.
Why It Matters
Chauvenet's criterion appears in undergraduate physics and engineering lab courses, often alongside John Taylor's textbook on error analysis, as a first objective rule for discarding a bad reading. Many statisticians criticize it because the 1/2 cutoff is arbitrary. In formal work they prefer tests such as Grubbs' test, or reporting results both with and without the point.
Common Mistakes
Mistake: Applying the criterion again and again: rejecting a point, recomputing the mean and standard deviation, then testing the next most extreme value.
Correction: Repeated use shrinks s each round and can strip away legitimate data. Apply the criterion once, to the single most suspicious point, and record any rejection in your report.
