Can each outcome have different coefficients?

Yes; that is the usual arrangement.

Must outcomes have identical units?

No, but scaling and prior choices need care.

Does Bayesian mean the correlations are known?

No. They can themselves be uncertain and represented through a posterior distribution.

AI Glossary / AI and machine learning / Intermediate

Bayesian multivariate linear regression

Definition

Bayesian multivariate linear regression jointly models several numerical outcomes using linear relationships with predictors. It places probability distributions on unknown coefficients and on quantities such as residual covariance, then updates them with data to obtain posterior estimates and joint predictions.

What it means

Here multivariate refers to multiple outcomes. A model predicting both processing time and energy consumption is multivariate; a model predicting only energy from ten inputs has multiple predictors but only one outcome. This distinction matters because terminology is sometimes used loosely.
Joint modeling can represent how outcomes move together after the predictors have been taken into account. Two production measures might remain correlated because of unmeasured operating conditions. Representing that residual relationship can matter when a decision depends on both measures at once.
Bayesian inference adds uncertainty about the relationships and their parameters. It does not establish that one outcome causes the other.

How it works

Arrange the observations as a table with one column per outcome, and use a predictor table to explain their means. A coefficient matrix gives each outcome its own regression coefficients. In a common model, a multivariate normal error distribution describes residual variability and correlation across outcomes.
Choose priors for coefficients and the covariance structure, then infer their posterior distribution. Classical conjugate constructions use matrix-normal and inverse-Wishart distributions, although these are not the only choices. Other parameterizations separate standard deviations from correlations. Evaluate marginal predictions for each outcome and joint predictions for combinations that affect decisions. Check computation and prior sensitivity, especially when outcomes are numerous.

Three practical examples

A factory predicts both cycle time and electricity use for planned production runs.
An environmental researcher jointly models two sensor measurements while representing their residual association.
A service planner forecasts two related workload measures and estimates the chance that both exceed operational thresholds on the same day.

Think of it this way

Think of predicting several instruments in a small ensemble. Each instrument has its own part, but their deviations may occur together. Separate forecasts describe individual parts; a joint model also describes their coordination. The analogy stops at statistical dependence: correlated residuals do not imply that one instrument or outcome directs the others, and real data require a specified probability model.

Worked example

Suppose a model predicts two outcomes, A and B. For a particular input, assume their predicted means are 10 and 20. To isolate the role of dependence, temporarily treat the parameters as known. Let both residual variances equal 4 and their residual correlation equal 0.75.
The covariance is 0.75 times 2 times 2, or 3. The predicted sum has mean 30 and variance 4+4+2(3)=14. If the outcomes were incorrectly treated as independent, its variance would be only 8.
The individual variances are identical in both calculations, but uncertainty about their combined total differs. A full Bayesian prediction also averages over uncertainty in coefficients and covariance parameters. This example illustrates the joint structure, rather than calculating an entire posterior or demonstrating that multivariate modeling always improves accuracy.

Illustrative case study

Fictional case: Willow Manufacturing forecasts cooling-water use and electricity demand separately. Managers then add the two standardized demand indicators for planning, assuming their errors are independent. During certain production runs, both demands rise together.
An analyst fits a Bayesian multivariate regression with product mix and run length as predictors, and checks whether residual association remains. The review compares joint predictions with a simpler separate-model baseline and inspects performance on later runs. Management receives an explanation of when combined uncertainty changes materially. No numerical improvement is asserted: the case illustrates a reason to assess dependence, not a guarantee that extra model complexity pays off.

Common misconceptions and mistakes

Multiple predictors do not by themselves make a model multivariate in the outcome sense used here. Residual correlation is not the same as raw correlation: predictors may explain part of the observed relationship.
Do not choose a covariance prior solely because its name sounds standard. Some priors couple variances and correlations in consequential ways. Missing outcomes require a deliberate modeling or data-handling strategy. A joint model can be poorly estimated when data are scarce relative to the number of covariance parameters, and it need not outperform separate regressions.

Business use

The method is useful when decisions depend on several continuous quantities together, such as combined resource demand or trade-offs between operating measures. Its practical value often lies in joint uncertainty rather than a more impressive point estimate.
Teams should identify the particular joint question before adding outcomes. Record measurement units, transformations, missing-data assumptions and covariance priors. Check whether changing correlations would alter the decision. Explain that predicted relationships may fail under unfamiliar operating conditions. If the business only needs one outcome, or residual dependence adds little actionable information, a simpler model can be easier to validate and maintain.

Check your understanding

A model predicts sales from price, season and advertising, but predicts no other outcome. Is it multivariate regression in this entry's sense?

Answer: No. It has multiple predictors and one outcome. Adding a jointly modeled second outcome, such as returns, would introduce the multiple-outcome structure discussed here.

Related terms

Learn next

Learn covariance matrices next, then compare separate and joint forecasts using a decision that depends on both outcomes. Explore priors that treat scales and correlations separately, and check whether joint uncertainty is calibrated on observations excluded from model fitting.

Sources

Stan User’s Guide, Multivariate outcomes, continuous-outcome covariance model and separate scale/correlation priors, consulted: https://mc-stan.org/docs/2_28/stan-users-guide/multivariate-outcomes.html. Sinay and Hsu, Bayesian Inference of a Multivariate Regression Model, introduction and Section 2 consulted via the primary-paper copy: https://pdfs.semanticscholar.org/3d41/cb10a71df7aeb2d47b79ef65bbf9731a856a.pdf. Wiley access failed; the accessible paper supplied the substantive review. The covariance example is original.

Previous: Bayesian model reduction Next: Bayesian network