Is this a test of time-series autocorrelation?

No; that is a different Durbin-named procedure.

Can any table with missing cells be analyzed this way?

No. The balanced incomplete design conditions must be checked.

Does rejection tell us which treatments differ?

Not by itself; it is an overall test.

AI Glossary / Statistics and probability / Advanced

Durbin test

Definition

The Durbin test is a rank-based test of treatment effects in a balanced incomplete block design. Each block receives only a subset of treatments, but the allocation is balanced so that treatments occur equally often and every treatment pair occurs together equally often. The test uses ranks within blocks to assess whether treatment effects differ. In a complete block design it reduces to the Friedman test. This meaning is distinct from the Durbin-Watson serial-correlation statistic and Durbin tests of econometric endogeneity.

What it means

Blocking groups observations that share relevant conditions, such as a testing session or production batch. Sometimes a block cannot accommodate every treatment, making an incomplete design necessary. Balance provides a structured way to compare treatments despite those planned omissions. Durbin's test focuses on relative ordering within each block rather than assuming the raw measurements follow a normal error distribution. It still requires a suitable experimental design and meaningful ranks; nonparametric does not mean that arbitrary missingness, dependence or allocation can be ignored.

How it works

Verify the incomplete-block incidence structure before testing. Rank observed responses separately inside each block and sum ranks for each treatment. With t treatments, b blocks, k observations per block and r occurrences per treatment, define A as the sum of all squared within-block ranks and C = bk(k + 1)^2/4. A common statistic is T = (t - 1)[sum of squared treatment rank sums - rC]/(A - C). Its usual large-sample reference is chi-square with t - 1 degrees of freedom. Implementations may use other approximations, and tie treatment must be checked. Independent blocks and an appropriate null randomization or exchangeability structure are important for interpretation.

Three practical examples

Product evaluation: compare several formulations when each independent tasting block can assess only a balanced subset.
Production experiments: use a planned incomplete allocation when a batch cannot accommodate every processing treatment.
Analysis auditing: verify that software receiving missing cells is actually being given a balanced incomplete design.

Think of it this way

Think of a tournament in which not every competitor appears in every round, but each competitor plays equally often and each pair meets equally often. Ranking within rounds can then support a structured overall comparison. An arbitrary set of abandoned matches would not automatically preserve that design.

Worked example

Consider three treatments A, B and C in three independent blocks containing AB, AC and BC. Each treatment appears twice and each pair once. Suppose the within-block ranks are A1/B2, A1/C2 and B1/C2. Treatment rank sums are two, three and four, whose squared sum is 29. Here A = 15, C = 3 x 2 x 9/4 = 13.5 and rC = 27, giving T = 2(29 - 27)/(15 - 13.5) = 8/3. A chi-square reference with two degrees of freedom gives approximately 0.2636. However, under independent equally likely rank assignments within these tiny two-observation blocks, enumeration gives eight assignments, six with statistic at least 8/3, yielding an exact randomization tail probability of 0.75. This original example shows why an asymptotic approximation can be poor in a very small design; the two probabilities are not interchangeable.

Illustrative case study

Fictional case: Cedar Product Lab plans a balanced subset of prototypes for each independent evaluation session. Before analysis, a statistician checks treatment occurrence and pair counts, then discovers that canceled sessions have broken the original balance. The team revisits the appropriate analysis rather than submitting the incomplete table unchanged to a Durbin routine. This invented case emphasizes design verification and honest treatment of missing observations. It does not claim that every disrupted experiment is unusable or that a rank method automatically fixes its imbalance.

Common misconceptions and mistakes

Do not confuse planned incomplete blocks with uncontrolled missing responses. Some software does not check whether the input is a true balanced incomplete block design, and tie handling varies. Rank all observations within their own blocks, not across the entire dataset. A significant omnibus result does not identify every differing treatment pair; follow-up comparisons require appropriate procedures and multiplicity handling. Differences in distributional shape can complicate interpretation as a simple location effect. A non-significant result does not establish equal effects or practical equivalence.

Business use

Businesses can use this framework when resource or fatigue constraints limit the number of alternatives evaluated in each block. The main value begins at design time: balanced allocation supports interpretable comparisons later. Preserve treatment, block and outcome identifiers, and record cancellations and ties. Report the actual reference distribution or exact procedure used, particularly in small studies. Decisions should incorporate effect descriptions and operational importance as well as a global p-value, because ranks convey relative performance without retaining every feature of the original measurement scale.

Check your understanding

Why is the three-block example balanced?

Answer: Every treatment appears in two blocks and every pair occurs together once, with two treatments in each block. Those incidence properties concern the design, not the observed ranks.

Related terms

Friedman test
Treatment effect
balanced incomplete block design
within-block rank
randomization test
chi-square approximation

Learn next

verify treatment-incidence balance from a design table and compare exact randomization calculations with large-sample approximations before choosing a reporting method.

Sources

Sources: NIST Dataplot, Durbin Test, design and rank-statistic definitions: https://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/durbin.htm ; PMCMRplus authors, durbinTest documentation, chi-square reference and implementation warnings: https://search.r-project.org/CRAN/refmans/PMCMRplus/html/durbinTest.html .

Previous: Duration Next: Durbin-Watson statistic