Can gamma be calculated when all values on one variable are equal?

No; every pair is tied on that variable, making the denominator zero.

Does gamma require equally spaced categories?

No. It requires meaningful ordering.

Does gamma equal one imply no ties?

No. It means there are no discordant pairs among the comparable pairs.

AI Glossary / Statistics and probability / Intermediate

Goodman and Kruskal's gamma

Definition

Goodman and Kruskal's gamma is a measure of ordinal association based on concordant and discordant pairs of observations. It is calculated as gamma = (C - D)/(C + D), where C counts concordant pairs and D counts discordant pairs. Pairs tied on either variable are excluded from this denominator. Gamma ranges from -1 to +1 when defined, describing the direction and balance of ordering among the comparable pairs.

What it means

Suppose each customer provides an ordered satisfaction rating and an ordered likelihood-to-recommend rating. Gamma asks whether customers ranked higher on one measure also tend to rank higher on the other. It uses order, not the assumed numerical distance between rating categories.

A positive value indicates more concordant than discordant pairs; a negative value indicates the reverse. Zero means equal counts among those two types of pair, not proof that the variables are independent.

Ties are central to interpretation. A large gamma can describe strong ordering among a relatively small subset of all pairs, while many pairs provide no comparison because at least one rating is equal. The coefficient should therefore be reported alongside information about ties and the underlying table.

How it works

For each distinct pair of records, compare their values on both ordered variables. If the record with the higher first value also has the higher second value, count a concordant pair. If it has the lower second value, count a discordant pair. If either variable is equal, count neither.

Calculate gamma from C and D. It is undefined when C + D = 0, because no pair supports the required comparison. Swapping the two variables leaves gamma unchanged. Reversing the order of exactly one variable changes its sign; strictly increasing recoding that preserves ordering and ties leaves it unchanged.

For larger datasets, counts can be obtained from an ordered contingency table without listing every pair. Inference must use a suitable variance estimate or resampling design. The many pairs are not independent observations because they repeatedly reuse the same records.

Three practical examples

Customer research: summarize whether ordered satisfaction and recommendation categories tend to increase together, while documenting how frequently responses are tied.
AI evaluation: compare ordered human quality ratings with ordered model-score bands, using gamma as one association summary rather than a measure of exact rating agreement.
Operations analysis: examine whether ordered service-complexity levels align with ordered delay categories, without interpreting the association as a causal effect of complexity.

Think of it this way

Think of comparing two ordered lists that allow shared positions. For each pair that has a clear ordering in both lists, ask whether the two lists agree about who comes first. Gamma balances agreements against disagreements and sets tied comparisons aside. It does not tell you how far apart the positions are or how many comparisons were set aside.

Worked example

Consider five original paired ordinal observations: A = (1,1), B = (1,3), C = (2,2), D = (3,2), and E = (3,3). There are choose(5,2) = 10 record pairs.

The concordant pairs are A - C, A - D, A - E, and C - E, so C = 4. The discordant pairs are B - C and B - D, so D = 2. The four remaining pairs are tied on at least one variable: A - B, B - E, C - D, and D - E. Hence gamma = (4 - 2)/(4 + 2) = 1/3, approximately 0.3333.

Independent enumeration using the sign of (x_i - x_j)(y_i - y_j) confirms four positive products, two negative products, and four zeros. Among the six comparable pairs, the concordant fraction is 4/6 = 2/3. Gamma equals twice that fraction minus one, giving the same result.

Now change only B to (1,2). The pair counts become five concordant, zero discordant, and five tied. Gamma becomes one, although half of all pairs are excluded. This does not mean the two variables have identical values or that every record has a unique ranking. The change illustrates why a coefficient of one needs its tie context. These five records are a calculation example, not a reliable population estimate.

Illustrative case study

In a fictional training platform, analysts compare learners' ordered confidence ratings with ordered assessment-performance bands. They calculate gamma and inspect the full table to see which combinations occur and how often ties arise.

A positive gamma suggests that higher confidence generally accompanies higher performance among comparable pairs. However, some learners report high confidence despite low scores. The team keeps those cases visible and evaluates any proposed feedback intervention separately. The fictional analysis illustrates an association summary; it does not show that confidence causes performance or that the platform can reliably classify individual learners from confidence alone.

Common misconceptions and mistakes

Common mistakes include treating gamma as a percentage of all pairs that agree, ignoring ties, applying it to categories with no meaningful order, and assuming a positive value establishes causality.

Gamma is also not interchangeable with Pearson correlation, Kendall's tau-b, or Somers' D. They use different information or denominators and can differ substantially when ties are common. A high association does not establish exact agreement: two raters may order cases similarly while one systematically assigns higher categories.

Do not assume a large number of pair comparisons creates an equally large independent sample. Sampling uncertainty depends on the original observations and their dependence structure, including repeated ratings or clustered respondents.

Business use

Use gamma when both variables are ordered and the exclusion of tied pairs suits the question. Record category order, pair counts, tie frequency, sample design, and uncertainty. For business decisions involving classification thresholds or exact agreement, add measures that answer those questions directly rather than relying on ordinal association alone.

Check your understanding

Answer: A dataset has 30 concordant pairs, 10 discordant pairs, and 60 tied pairs.

What is gamma?

Answer: It is (30 - 10)/(30 + 10) = 0.5. The tied pairs do not enter the denominator, so interpreting 0.5 as a statement about half of all 100 pairs would be incorrect.

Related terms

Ordinal variable
Kendall's tau
Contingency table
concordance
discordance
tied ranks
Somers' D
Spearman correlation

Learn next

Learn pairwise ordering and contingency tables, then compare how ordinal association measures handle ties. Study agreement measures separately when the intended question concerns matching ratings rather than similar ordering.

Sources

SAS, PROC FREQ Measures of Association, Gamma section: https://support.sas.com/documentation/cdl/en/procstat/63963/HTML/default/procstat_freq_a0000000562.htm — primary documentation of ordinal comparisons, ties, and related measures. Wolfram, GoodmanKruskalGamma: https://reference.wolfram.com/language/ref/GoodmanKruskalGamma.html.en — primary implementation definition based on concordant and discordant counts. All five-record examples and pair enumerations above are original.

Previous: Gompertz distribution Next: Goodness-of-fit