Must all multivariate models be normal?

No; the worked example is discrete.

AI Glossary / Statistics and probability / Advanced

Multivariate model

Definition

A multivariate model simultaneously describes or predicts more than one response variable. The responses can be different outcomes or measurements of an outcome at different times.

What it means

The word concerns the responses being modeled. A multivariable model has several predictors; a multivariate model has several responses. A model can be both. For example, predicting two service-time measures from request length and queue size is both multivariate and multivariable.

How it works

Specify the response vector, its relationship with predictors, and the joint variation relevant to the analysis. Depending on the model, this may involve response-specific equations and a covariance structure. The details depend on whether the responses are continuous, categorical, repeated over time, or otherwise structured.
The practical question is what the joint model needs to support: separate predictions, combinations of outcomes, simultaneous uncertainty statements, or another clearly defined task.

Three practical examples

A fictional service team predicts preparation time and execution time for the same request, then investigates their combined duration.
A fictional training project models two assessment scores from each participant rather than replacing them with an arbitrary single score.
A fictional product study predicts a user's activity at two later checkpoints, preserving the connection between repeated measurements from the same user.

Think of it this way

Two instruments playing together can each have a recognizable part, while their coordination changes the overall sound. Describing each part separately does not necessarily describe what happens when they are combined. A joint model makes the relevant relationship explicit.

Worked example

Consider an original service-time illustration with two responses, Y1 and Y2, measured in the same time unit. Let x be a numerical workload measure. Suppose the conditional mean equations are E(Y1 given x)=10+2x and E(Y2 given x)=5+x.
At x=2, the predicted means are 14 and 7. The predicted mean of total time T=Y1+Y2 is therefore 21. The predicted mean of the difference D=Y1-Y2 is seven.
These are conditional mean predictions, not promises about individual requests. To study variation around them, write each response as its mean plus an error: Y1=14+e1 and Y2=7+e2 at this workload.
Suppose the errors have variances four and nine, and covariance three. Their standard deviations are two and three, giving correlation 3/(2 times 3)=0.5. The covariance matrix has diagonal entries four and nine and off-diagonal entries three.
The total-time variance is four plus nine plus twice three, or 19. Its standard deviation is sqrt(19), approximately 4.36. The difference variance is four plus nine minus twice three, or seven, with standard deviation approximately 2.65.
If an analyst incorrectly assumed zero covariance, both variances would be reported as 13, with standard deviation approximately 3.61. That assumption would understate variability in total time and overstate variability in the difference. The individual response means and variances would not reveal this error by themselves.
Here is a fully specified invented error model that produces the stated covariance. Independently draw U and V, each taking values minus one and plus one with equal probability. Set e1=2U and e2=1.5U+(3 sqrt(3)/2)V.
Each error has mean zero. The first variance is four. The second is 1.5 squared plus (3 sqrt(3)/2) squared, which is 2.25+6.75=9. Their covariance is 2 times 1.5 times Var(U)=3, because the contribution involving U times V has expectation zero.
There are four equally likely error pairs. Rounded to three decimals, they are (2, 4.098), (2, -1.098), (-2, 1.098), and (-2, -4.098). Adding these to the predicted means produces the four possible response pairs in this deliberately simple model.
The resulting total times are approximately 27.098, 21.902, 20.098, and 14.902. For a total-time threshold of 25, only the first outcome exceeds the threshold, so the probability is one quarter.
That probability required the explicitly specified joint distribution. Means and a covariance matrix alone would not determine it. In particular, inserting the same means and covariance into a normal model would describe a different distribution; the four-point construction is not normal.
For comparison, change the second error equation to e2=-1.5U+(3 sqrt(3)/2)V. Its mean and variance remain zero and nine, but its covariance with e1 becomes minus three. Total-time variance now becomes seven, while difference variance becomes 19. The same individual means and variances therefore support substantially different joint behavior.
Finally, add another predictor, z, to both mean equations, such as an indicator for a request type. That changes the model to have multiple predictors as well as multiple responses. By contrast, predicting only total time from x and z would be a multivariable model with one response. The number of predictors does not settle whether a model is multivariate.

Illustrative case study

Fictional company Birch Services initially reports preparation and execution forecasts separately. When managers request a forecast for their sum, analysts discover that the two errors tend to move together.
The team documents the responses, units, and dependence assumptions, then evaluates the combined forecast on later requests. Its presentation distinguishes the uncertainty of each component from uncertainty in the total. This fictional case illustrates why the desired decision should guide the joint analysis.

Common misconceptions and mistakes

Many predictors do not automatically make a model multivariate.
Two accurate-looking component means do not establish accurate joint predictions.
Zero covariance is an assumption to justify, not a default implied by separate output columns.
A covariance matrix alone does not specify all probabilities.
Joint modeling does not automatically improve every prediction.

Business use

Define the outcome combination that matters before choosing a model. In the fictional service example, a staffing decision based on total duration needs evidence about combined errors, not only separate accuracy summaries for preparation and execution.

Check your understanding

What is the predicted total at x=2?

Answer: Twenty-one.

What is its variance with covariance three?

Answer: Nineteen.

Why is the threshold probability one quarter?

Answer: Exactly one of the four equally likely outcomes exceeds 25.

Related terms

Multivariable model
Covariance
Multivariate regression
response vector
joint distribution
repeated measures
multi-output prediction

Learn next

Compare marginal and joint predictions, then study multivariate regression and models for repeated observations.

Multivariate regression

Sources

Frank Harrell, Glossary of Statistical Terms, multivariate model and multivariable model: https://hbiostat.org/glossary/ . Equations, finite error distributions, calculations, and company scenario are original.

Previous: Multivariate Laplace distribution Next: Multivariate normal distribution