Can useful contrasts be nonorthogonal?

Yes; substantive questions need not be perpendicular.

How many mutually orthogonal contrasts are possible for k means?

At most k minus one nonzero contrasts.

Does multiplying a contrast by a constant change its question?

It changes units and direction if negative, but a nonzero scaling preserves the corresponding zero-null hypothesis.

AI Glossary / Statistics and probability / Intermediate

Orthogonal contrasts

Definition

Orthogonal contrasts are linear comparisons of group means that are uncorrelated under a specified statistical model. In a balanced one-way design with independent observations and common variance, two contrasts are orthogonal when the sum of the products of their corresponding coefficients is zero. Each contrast's coefficients must separately sum to zero.

What it means

A contrast expresses a focused question about groups, such as whether two training methods differ or whether their average differs from a control. Orthogonal contrasts separate distinct directions of comparison in the group means.
The statistical meaning depends on the design and covariance structure. Calling coefficient vectors perpendicular is not enough when group sizes differ or observations are correlated. Nor does orthogonality automatically make every associated significance test independent, particularly when tests share an estimated error variance.

How it works

Write a comparison as L = sum of c_i times mu_i, where mu_i is a population group mean and the coefficients c_i sum to zero. Estimate it by replacing population means with sample means.
For two contrasts with coefficient vectors c and d, independent groups, common observation variance sigma squared, and group sizes n_i, the covariance of their estimates is sigma squared times the sum of c_i d_i/n_i. A zero value establishes uncorrelated estimates under that model.
When all n_i are equal, the condition simplifies to the ordinary coefficient dot product being zero. With unequal group sizes, that simplification can fail. With more general models, use the covariance matrix of estimated means: the appropriate condition is c-transpose times V times d equals zero.
Under joint normality, uncorrelated contrast estimates are independent. Without joint normality, zero covariance does not generally imply independence. Test statistics that divide estimates by a shared random variance estimate require additional care.
A complete set for k groups contains at most k minus one mutually orthogonal, nonzero contrasts. In the usual balanced one-way analysis, a complete set partitions the between-group sum of squares into contrast components.
Choose comparisons for substantive reasons, preferably before examining results. Orthogonal comparisons can be useful, but forcing a scientifically important question into an arbitrary orthogonal scheme is not necessarily helpful. Specify the intended error-rate control when making multiple inferential claims.

Three practical examples

Training evaluation: compare two new teaching methods with each other, and separately compare their average with a control group in a balanced study.
Product research: compare two interface variants and distinguish that comparison from the overall contrast between redesigned and existing interfaces.
Process experimentation: separate a linear trend across equally spaced settings from a quadratic trend using appropriate planned contrast coefficients.

Think of it this way

Think of measuring movement east-west and north-south on a map. These directions separate different components of movement. But the statistical equivalent also depends on how uncertain the measurements are; unequal precision changes the geometry relevant to the comparison.

Worked example

Suppose three independent fictional training groups A, B, and control each contain ten participants. Their mean scores are 72, 68, and 60. Assume common observation variance for the planned analysis.
Contrast C compares A with B using coefficients (1, -1, 0). Its estimate is 72 - 68 = 4. Contrast D compares their average with control using (0.5, 0.5, -1). Its estimate is 0.5 times 72 + 0.5 times 68 - 60 = 10.
Both coefficient sets sum to zero. Their dot product is 1 times 0.5 + (-1) times 0.5 + 0 times (-1) = 0. Because sample sizes are equal, the estimates have zero covariance under the stated model. Their variances are 0.2 sigma squared and 0.15 sigma squared, respectively.
Now change the sample sizes to 5, 10, and 10 without changing the coefficient vectors. The covariance multiplier becomes 0.5/5 - 0.5/10 = 0.05, which is not zero. The same written comparisons are no longer statistically orthogonal under the common-variance model.
The mean differences alone do not determine significance; an error estimate and valid sampling assumptions are still needed.

Illustrative case study

Illustrative fictional case: Birch Learning plans a three-arm course trial. Before collecting data, its analysts specify the two comparisons in the example and aim for equal allocation.
When withdrawals leave unequal group sizes, they retain the substantive questions but calculate their covariance using the actual analysis model. They report the comparisons with suitable uncertainty rather than describing them as automatically independent. This fictional case illustrates how a planned design property can change when the final data differ from the plan.

Common misconceptions and mistakes

Do not confuse coefficients summing to zero with products of two coefficient sets summing to zero: the first defines a contrast; the second is the balanced-design orthogonality condition.
Do not assume zero covariance always means independence. Avoid ignoring unequal sample sizes, repeated observations, or unequal variances. Orthogonality also does not eliminate multiplicity concerns or make data-selected comparisons equivalent to preplanned hypotheses.

Business use

Orthogonal contrasts help organizations organize clear questions within a multi-group experiment. Their value is interpretability and a transparent relationship among comparisons. Reports should name each comparison, state the weighting of groups, describe the covariance assumptions, and explain how uncertainty and multiple claims were handled.

Check your understanding

Why are the example's first two contrasts orthogonal with equal sample sizes?

Answer: Their coefficient product sum is zero under the independent, common-variance model.

Why does unequal allocation matter?

Answer: The covariance uses inverse sample-size weights, so the ordinary dot product is no longer sufficient.

Related terms

Analysis of variance
Covariance
Multiple comparisons
Linear contrast
planned comparison
balanced design
polynomial contrast
estimated marginal means

Learn next

Next study contrast standard errors and confidence intervals. Then examine how unbalanced designs and repeated measurements change the covariance matrix used to evaluate comparisons.

Covariance matrix

Sources

Penn State, STAT 503, Experiments with a Single Factor: https://online.stat.psu.edu/stat503/Lesson03 — primary search excerpts on contrast orthogonality and unequal sample sizes; full page retrieval failed. University of Vermont, Statistical Methods glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html — terminology context. The covariance derivation, worked calculations, and fictional case are original.

Previous: Orthogonal array testing Next: Orthogonal Matching Pursuit