AI Glossary / Statistics and probability / Advanced

Multivariate procedures

Definition

Multivariate procedures are statistical methods that analyze two or more response variables simultaneously. In broader statistical usage, the term also covers methods for exploring the structure of several variables together.

What it means

The supplied glossary uses the multiple-response meaning. Examples include multivariate analysis of variance and multivariate regression. The wider field also includes principal component analysis, factor analysis, and canonical correlation. These methods answer different questions; they are not interchangeable merely because each uses several variables.

How it works

Start with the question, identify the variables and observational units, and choose a procedure appropriate to their joint structure. Specify the hypothesis or prediction target and the assumptions used to quantify uncertainty. Examine the results jointly and then interpret individual outcomes with an appropriate follow-up plan.
A multivariate analysis can use covariance information that separate summaries overlook. It does not automatically improve every analysis or justify every conclusion about each component variable.

Three practical examples

A fictional training study compares two assessment outcomes between teaching formats using a planned joint question.
A fictional service study predicts preparation and execution durations together, then evaluates combinations relevant to scheduling.
A fictional survey project uses a dimension-reduction method to summarize related measurements, recognizing that summarization is a different aim from testing group differences.

Think of it this way

A dashboard can show several indicators separately or examine how they move together. A multivariate procedure supplies a defined way to answer a joint question. The choice of method matters as much as the decision to look at several indicators.

Worked example

Consider an original teaching example with two continuous response scores measured on each of 25 independent units. Assume their joint distribution is normal and that their population covariance is known, with rows (25, 20) and (20, 25).
The variances of the two sample means are each 25/25=1. Their covariance is 20/25=0.8. Thus the covariance matrix V of the sample-mean vector has rows (1, 0.8) and (0.8, 1).
The null hypothesis states that the population mean vector equals a prespecified reference vector. Suppose the observed mean differences from that reference are d=(1, -1).
Looking at either outcome separately, its mean difference is only one standard error from zero. The corresponding two-sided standard-normal probability is about 0.3173 for each outcome.
For a joint test, calculate Q=d-transpose times inverse(V) times d. The inverse matrix is one divided by 0.36 times the matrix with rows (1, -0.8) and (-0.8, 1).
Substitution gives Q=(1+1.6+1)/0.36=10. Under the stated null hypothesis, known covariance, independent units, and normal model, Q has a chi-square distribution with two degrees of freedom. Its upper-tail probability is exp(-10/2), approximately 0.006738.
The joint result differs from the separate results because the observed mean shifts point in opposite directions while the reference errors are positively related. That particular combination is relatively unusual under the specified null model.
Now change only the observed difference vector to d=(1, 1). The individual standardized mean differences still have magnitude one. The joint statistic, however, becomes (1-1.6+1)/0.36=10/9, approximately 1.111.
Its upper-tail probability is exp(-5/9), approximately 0.57375. Moving together is much more consistent with the positive covariance pattern in this example. Reporting only two separate standardized differences would conceal the distinction between these two joint results.
The contrast between the outcomes gives another view. For d=(1, -1), the estimated mean contrast is one minus minus one, or two. Its variance is one plus one minus twice 0.8, or 0.4. Its standardized value is therefore 2/sqrt(0.4), approximately 3.162.
That contrast targets a particular direction; the joint test targets departures in any direction. Their probabilities need not be identical. A contrast chosen after inspecting the results also raises selection issues that a prespecified contrast avoids.
The example uses a known covariance to keep the arithmetic transparent. In real work, covariance is usually estimated. The resulting statistic, reference distribution, and assumptions must match the selected procedure; the known-covariance chi-square calculation should not simply be reused without justification.
The number of units is 25, not 50. Measuring two outcomes on the same unit supplies a response pair, not two independent units. Treating the two columns as independent extra observations would change the uncertainty calculation incorrectly.
Suppose the second score is converted to units 100 times larger. Its mean difference and standard deviation are multiplied by 100, and the covariance matrix is transformed consistently. The joint quadratic statistic remains unchanged. An analysis that changes only the recorded scores but leaves the covariance unchanged is inconsistent.
Finally, suppose the second outcome is defined to be exactly twice the first. The two columns then contain redundant information, and their covariance matrix is singular. An ordinary matrix inverse is unavailable. The analyst must recognize the redundancy and choose an appropriate representation or method rather than treating the second column as an independent source of evidence.

Illustrative case study

Fictional company Alder Learning compares two assessment outcomes from the same participants. Before analyzing results, it states a joint question, identifies the independent units, and specifies how any follow-up comparisons will be handled.
The report includes the outcome-specific estimates as well as the joint result. It explains that rejecting a joint null does not establish that both outcomes changed individually. This fictional case illustrates planned interpretation rather than a promise that a joint test will find an effect.

Common misconceptions and mistakes

Several predictors alone do not make an analysis a multiple-response procedure.
A significant joint test does not prove every response differs individually.
Repeated measurements on one unit are not automatically independent observations.
Different multivariate methods solve different problems.
A matrix calculation does not remove the need to check assumptions and interpret the study design.

Business use

Choose the procedure around the decision. In the fictional training study, a joint question can help assess an outcome profile, while practical decisions still require the size and meaning of each change. Distinguish exploratory summaries, predictive performance, and confirmatory tests in the report.

Check your understanding

What is Q for differences (1, -1)?

Answer: Ten.

What is Q for differences (1, 1)?

Answer: Ten ninths.

How many independent units are in the example?

Answer: Twenty-five.

Related terms

Learn next

Compare joint mean tests with planned contrasts, then study the assumptions of the multivariate procedure suited to your question.

Sources

David Howell, Statistical Methods for Psychology glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html . Statsmodels, Multivariate Statistics: https://www.statsmodels.org/stable/multivariate.html and MANOVA: https://www.statsmodels.org/stable/generated/statsmodels.multivariate.manova.MANOVA.html . Numerical illustration and company scenario are original.

Previous: Multivariate outliers Next: Multivariate random variable