Is Q = 0.8 an eighty-percent probability of an error?

No. It is a ratio of gaps; a p-value requires the appropriate null reference, and neither quantity is directly an error probability.

Can a significant result prove a transcription mistake?

No. It indicates incompatibility with the assumed model, not the cause.

Are all Dixon tests the same formula?

No. Several variants and testing conventions exist.

AI Glossary / Statistics and probability / Intermediate

Dixon's Q test

Definition

Dixon's Q test is an outlier test based on ratios of gaps between ordered observations. Its familiar small-sample form compares the gap next to a suspected extreme value with the full sample range. A sufficiently large ratio can indicate that the extreme observation is inconsistent with the assumed common normal population. Several Dixon statistics and tail conventions exist, so the formula, sample size and reference distribution must match. A flagged value requires investigation rather than automatic deletion.

What it means

The test asks whether an extreme observation is unusually separated from its neighbors under a particular model of otherwise comparable measurements. It is often associated with small laboratory datasets. The question is not whether a value is inconvenient, above a business limit or simply far from the mean. A valid unusual observation can reveal a meaningful process change. The test's assumptions and intended single-outlier setting must therefore be considered alongside the measurement history and the consequences of excluding data.

How it works

Order the observations from x_(1) to x_(n) and identify the planned extreme-value comparison. For the simple upper-end r10 form, Q = [x_(n) - x_(n-1)]/[x_(n) - x_(1)]; the lower-end version uses the gap between the first two values. Compare with critical values or a p-value for the same statistic, sample size and one- or two-sided procedure. Refined Dixon variants use different gaps and denominators for different sample-size ranges. Check the assumption of a single approximately normal population and relevant independence. If all values coincide, the range is zero and the ratio is undefined. Preserve the original observations and document any follow-up decision.

Three practical examples

Laboratory review: investigate a single extreme result among a small set of comparable replicate measurements.
Instrument validation: assess whether an unusually separated reading warrants checking calibration, transcription or operating conditions.
Statistical training: compare gap-based outlier statistics with other procedures and study why reference tables cannot be interchanged.

Think of it this way

Think of people standing along a line and asking whether the person at one end is unusually isolated from the next person, relative to the group's total span. A large gap attracts attention, but it does not tell you whether someone stood in the wrong place or whether the arrangement reflects something real.

Worked example

Suppose five hypothetical replicate measurements, already ordered, are 10.0, 10.1, 10.2, 10.3 and 11.5 units. For the simple upper-end r10 statistic, the adjacent gap is 11.5 - 10.3 = 1.2 and the range is 11.5 - 10.0 = 1.5. Thus Q = 1.2/1.5 = 0.8. This is the statistic, not a p-value. A formal decision requires the matching five-observation reference and the specified testing convention. If every reading is converted using y = 1000x + 7, both gap and range are multiplied by 1000, so Q remains 0.8. The invariance shows that changing measurement units and origin does not alter this ratio. It does not establish that the high reading is erroneous or justify choosing a reference table after seeing the result.

Illustrative case study

Fictional case: Alder Measurement Services notices a high reading in a small replicate series. Analysts calculate the prespecified Dixon statistic, verify the correct reference procedure and examine the instrument log. They preserve the raw result while investigating whether the reading reflects a preparation error, a measurement problem or genuine variation. Any exclusion is documented with its rationale and the analysis is examined with and without the observation. This invented case illustrates disciplined outlier review; it does not claim that every extreme result should be removed or that a statistical flag supplies a physical explanation.

Common misconceptions and mistakes

Do not use Q as a universal cutoff independent of sample size. Do not confuse a ratio statistic with its p-value or mix critical values for different Dixon variants. Multiple outliers can mask one another, and repeated testing after successive deletions changes the overall error behavior. Strong skewness, mixed populations or dependent readings can invalidate the normal single-population reference. An outlier test does not detect every bad measurement and cannot distinguish a valid rare event from an error by itself. A specification-limit violation is a different question from statistical outlier status.

Business use

Businesses should use outlier tests as part of a documented measurement-quality process. The main value is a reproducible signal for investigation when the method's scope fits the data. For AI training datasets, indiscriminate deletion of extreme observations can remove precisely the rare cases a model must handle. Review provenance, operational context and sensitivity of conclusions before excluding values. When multiple unusual observations or complex distributions are expected, consider methods designed for that setting rather than repeatedly applying a small-sample single-outlier procedure.

Check your understanding

Why must the sample-size-specific reference be consulted after calculating Q?

Answer: The distribution of extreme gaps changes with the number of observations and the chosen statistic. The same ratio can have different significance under different procedures, so the number 0.8 alone cannot determine a valid testing conclusion.

Related terms

Outlier
Order statistic
Masking
Robust statistics
Measurement error
sample range
gap statistic
normality assumption
multiple testing
Grubbs' test

Learn next

calculate upper and lower gap ratios for a small dataset and identify the exact test variant. Study how multiple extremes affect outlier detection and why sequential deletion changes error rates. Then separate statistical flags, measurement investigations and justified data-exclusion decisions in a reproducible analysis record.

Sources

NIST Dataplot, Dixon Test: https://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/dixon.htm NIST/SEMATECH, Detection of Outliers: https://itl.nist.gov/div898/handbook/eda/section3/eda35h.htm dixonTest package authors, Dixons Outlier Test: https://search.r-project.org/CRAN/refmans/dixonTest/html/dixonTest.html

Previous: Divisive Communities Next: DMAIC