Are multiple and multivariate regression really different? What is a variate anyways?


Very quickly, I would say: 'multiple' applies to the number of predictors that enter the model (or equivalently the design matrix) with a single outcome (Y response), while 'multivariate' refers to a matrix of response vectors. Cannot remember the author who starts its introductory section on multivariate modeling with that consideration, but I think it is Brian Everitt in his textbook An R and S-Plus Companion to Multivariate Analysis. For a thorough discussion about this, I would suggest to look at his latest book, Multivariable Modeling and Multivariate Analysis for the Behavioral Sciences.

For 'variate', I would say this is a common way to refer to any random variable that follows a known or hypothesized distribution, e.g. we speak of gaussian variates $X_i$ as a series of observations drawn from a normal distribution (with parameters $\mu$ and $\sigma^2$). In probabilistic terms, we said that these are some random realizations of X, with mathematical expectation $\mu$, and about 95% of them are expected to lie on the range $[\mu-2\sigma;\mu+2\sigma]$ .

  • 3
    $\begingroup$ Even coursera.org/learn/machine-learning/home/week/2 uses the term multivariate regression instead of multiple regression… $\endgroup$ Oct 28 '15 at 2:32
  • 2
    $\begingroup$ I think the same confusion arises with people using the term GLM for General Linear Model (e.g., in neuroimaging studies) vs. Generalised Linear Model. I have seen many instances of "multivariate logistic regression" where there's only one outcome, and I don't think this matters so much as long as the term is clearly defined by the author. $\endgroup$
    – chl
    Oct 28 '15 at 10:38

Here are two closely related examples which illustrate the ideas. The examples are somewhat US centric but the ideas can be extrapolated to other countries.

Example 1

Suppose that a university wishes to refine its admission criteria so that they admit 'better' students. Also, suppose that a student's grade Point Average (GPA) is what the university wishes to use as a performance metric for students. They have several criteria in mind such as high school GPA (HSGPA), SAT scores (SAT), Gender etc and would like to know which one of these criteria matter as far as GPA is concerned.

Solution: Multiple Regression

In the above context, there is one dependent variable (GPA) and you have multiple independent variables (HSGPA, SAT, Gender etc). You want to find out which one of the independent variables are good predictors for your dependent variable. You would use multiple regression to make this assessment.

Example 2

Instead of the above situation, suppose the admissions office wants to track student performance across time and wishes to determine which one of their criteria drives student performance across time. In other words, they have GPA scores for the four years that a student stays in school (say, GPA1, GPA2, GPA3, GPA4) and they want to know which one of the independent variables predict GPA scores better on a year-by-year basis. The admissions office hopes to find that the same independent variables predict performance across all four years so that their choice of admissions criteria ensures that student performance is consistently high across all four years.

Solution: Multivariate Regression

In example 2, we have multiple dependent variables (i.e., GPA1, GPA2, GPA3, GPA4) and multiple independent variables. In such a situation, you would use multivariate regression.

  • 5
    $\begingroup$ There's always one that properly answers the question with examples :) $\endgroup$ Aug 30 '17 at 2:44
  • 2
    $\begingroup$ 100% the best answer that you can actually understand $\endgroup$
    – Alvis
    Apr 16 '18 at 14:08

Simple regression pertains to one dependent variable ($y$) and one independent variable ($x$): $y = f(x)$

Multiple regression (aka multivariable regression) pertains to one dependent variable and multiple independent variables: $y = f(x_1, x_2, ..., x_n)$

Multivariate regression pertains to multiple dependent variables and multiple independent variables: $y_1, y_2, ..., y_m = f(x_1, x_2, ..., x_n)$. You may encounter problems where both the dependent and independent variables are arranged as matrices of variables (e.g. $y_{11}, y_{12}, ...$ and $x_{11}, x_{12}, ...$), so the expression may be written as $Y = f(X)$, where capital letters indicate matrices.

Further reading:

  • $\begingroup$ I understand the definition. But what is the effect of treating a multi-variate regression as a system of uni-variate regressions? $\endgroup$
    – LKS
    Aug 3 '16 at 19:30
  • $\begingroup$ @LKS: You may want to ask that in a completely separate question. $\endgroup$ Oct 5 '16 at 21:10
  • 2
    $\begingroup$ stats.stackexchange.com/questions/254254/… $\endgroup$
    – KarthikS
    Jan 26 '18 at 23:43
  • $\begingroup$ Did the answer in the Quora referring to this page? :P $\endgroup$
    – user167493
    Nov 6 '18 at 6:31

I think the key insight (and differentiator) here aside from the number of variables on either side of the equation is that for the case of multivariate regression, the goal is to utilize the fact that there is (generally) correlation between response variables (or outcomes). For example, in a medical trial, predictors might be weight, age, and race, and outcome variables are blood pressure and cholesterol. We could, in theory, create two "multiple regression" models, one regressing blood pressure on weight, age, and race, and a second model regressing cholesterol on those same factors. However, alternatively, we could create a single multivariate regression model that predicts both blood pressure and cholesterol simultaneously based on the three predictor variables. The idea being that the multivariate regression model may be better (more predictive) to the extent that it can learn more from the correlation between blood pressure and cholesterol in patients.

  • 1
    $\begingroup$ Great point. I was wondering if multivariate regression can be done with R. Using Manova, I am able to do multivariate ANOVA, but not able to get coefficients like univariate regression. $\endgroup$
    – KarthikS
    Jan 26 '18 at 23:45

In multivariate regression there are more than one dependent variable with different variances (or distributions). The predictor variables may be more than one or multiple. So it is may be a multiple regression with a matrix of dependent variables, i. e. multiple variances. But when we say multiple regression, we mean only one dependent variable with a single distribution or variance. The predictor variables are more than one. To summarise multiple refers to more than one predictor variables but multivariate refers to more than one dependent variables.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.