Must X and Y be independent?

No; rho describes their linear association.

Do normal marginals guarantee this model?

No; the joint structure matters.

Is the conditional variance the same as the marginal variance?

It equals the marginal variance times one minus rho squared.

AI Glossary / Statistics and probability / Advanced

Bivariate normal model

Definition

A bivariate normal model describes two random variables jointly using a two-dimensional normal distribution. It is specified by two means, two variances and their covariance, or equivalently their correlation and standard deviations. In regression settings, it models both variables as random and implies a normal conditional distribution with a linear conditional mean and constant conditional variance.

What it means

The model describes how paired values occur together, not merely whether each variable separately has a bell-shaped distribution. Two normal marginal distributions do not automatically make their joint distribution bivariate normal.
Treating both variables as random also does not mean that both are measured inaccurately. A measurement-error model needs an explicit description of the observation process beyond the bivariate normal distribution itself.

How it works

Write the means as mu_X and mu_Y, standard deviations as sigma_X and sigma_Y, and correlation as rho. The covariance is rho sigma_X sigma_Y. For positive standard deviations and absolute rho below one, the covariance matrix is positive definite and the joint density has elliptical contours.
Conditional on X = x, Y is normal with mean mu_Y + rho(sigma_Y/sigma_X)(x-mu_X) and variance sigma_Y squared times (1-rho squared). The conditional mean changes with x, but the conditional variance does not.
Under joint normality, zero correlation implies independence. This implication does not hold for arbitrary pairs of variables. Correlations of positive or negative one produce a degenerate distribution concentrated on a line rather than an ordinary two-dimensional density.

Three practical examples

Simulation design: researchers generate paired numeric measurements with specified means, scales and correlation to evaluate an analytical method.
Conditional forecasting: an analyst models a second numeric measurement after observing the first, checking whether linear conditional means and constant spread are plausible.
Statistical education: a team compares the marginal distribution of an outcome with its narrower conditional distribution and explains what information conditioning adds.

Think of it this way

Imagine a cloud of paired measurements on a chart. The whole cloud describes the joint distribution. A vertical slice at a particular X value describes the conditional distribution of Y. Its center can differ from the overall center even though it is part of the same cloud.

Worked example

Consider an original model for two simulated measurements. Let X have mean 10 and standard deviation 2. Let Y have mean 50 and standard deviation 5. Set their correlation to 0.6. Their covariance is 0.6 times 2 times 5 = 6, so the covariance matrix has rows (4, 6) and (6, 25).
The determinant is 4 times 25 minus 6 squared = 64, which is positive. Together with the positive diagonal variances, this confirms that the matrix defines a nondegenerate bivariate normal model.
Suppose X is observed to be 14, two standard deviations above its mean. The conditional mean of Y is 50 + 0.6(5/2)(14-10) = 56. The conditional variance is 25(1-0.36) = 16, giving standard deviation 4. Thus Y given X = 14 is normal with mean 56 and variance 16.
Using approximately 1.96 standard deviations on either side, a central conditional probability interval is 56 plus or minus 7.84, or approximately 48.16 to 63.84. Under this specified model it contains about 95 percent of the conditional distribution. It is not a confidence interval for the unknown conditional mean, and it does not include extra uncertainty from estimating parameters.
Before X is observed, Y has mean 50 and standard deviation 5. Conditioning has shifted the center to 56 and reduced the standard deviation to 4. The conditional-mean slope is 0.6 times 5/2 = 1.5, so its equation is E(Y given X = x) = 35 + 1.5x.
Now reverse the prediction question. Given Y = 60, the conditional mean of X is 10 + 0.6(2/5)(60-50) = 12.4. The conditional variance of X is 4(1-0.36) = 2.56, giving standard deviation 1.6. The reverse conditional-mean slope is 0.24, not the reciprocal of 1.5. Algebraically inverting 60 = 35 + 1.5x would instead give x approximately 16.6667, answering a different question.
The product of the two conditional-mean slopes is 1.5 times 0.24 = 0.36, equal to rho squared.
We can construct the same model from independent standard-normal variables Z and E. Set X = 10 + 2Z and Y = 50 + 3Z + 4E. Then Var(X) = 4, Var(Y) = 9 + 16 = 25 and Cov(X,Y) = 2 times 3 = 6. Observing X = 14 fixes Z = 2, leaving Y = 56 + 4E. This directly reproduces the conditional mean and variance above.

Illustrative case study

Fictional company Cedar Simulation creates paired numeric outputs to test a forecasting workflow. Its analysts record the stipulated joint distribution and compare theoretical conditional predictions with repeated simulated draws. They distinguish variation remaining in an individual outcome from uncertainty in a fitted mean.
Before applying the model to observed measurements, they inspect the paired data for curvature, unequal spread and unusual observations. They do not infer that normal-looking separate histograms prove the joint model.

Common misconceptions and mistakes

Normal marginals are insufficient for joint normality. For a counterexample, take a standard-normal Z and an independent random sign S that is equally likely to be positive or negative one. X = Z and Y = SZ are each normal and uncorrelated, but their absolute values are always equal. They are dependent and not jointly bivariate normal.
Another mistake is interpreting a conditional regression as a causal effect, or treating the reverse regression as its algebraic inverse. The model also permits values anywhere on the real line, so it can be unsuitable for bounded measurements or counts without an appropriate transformation or alternative specification.

Business use

The model provides a transparent baseline for paired continuous measurements and simulation studies. Reports should distinguish stipulated from estimated parameters, marginal from conditional uncertainty, and association from intervention effects.

Check your understanding

Why was the conditional standard deviation four?

Answer: The variance was 25 times 0.64 = 16.

Why was the reverse prediction 12.4?

Answer: It used the conditional distribution of X given Y.

What made the normal-marginal counterexample dependent?

Answer: The two absolute values were identical.

Related terms

Learn next

Vary rho in the simulation construction and compare the conditional spread, keeping both marginal standard deviations fixed.

Sources

Mikael Vejdemo-Johansson, CUNY Mathematical Statistics, Linear Regression, bivariate normal distribution and conditional regression formulas: https://www.math.csi.cuny.edu/~mvj/MTH410-S23/lectures/L20.html . The numerical model, counterexample and company scenario are original illustrations.

Previous: Bivariate analysis Next: Biword index