Joint Upper-Tail Probabilities and Correlation

Why Correlation Matters for Joint Exceedances

Suppose two measurements each have a known probability of exceeding a specified threshold. Those marginal probabilities alone do not determine the probability that both measurements exceed their thresholds. The relationship between the measurements matters: positive correlation tends to make high values occur together, while negative correlation tends to make them occur on opposite sides of their means.

For a bivariate normal pair, the joint upper-tail probability can be expressed through the bivariate normal cumulative distribution function. Keeping the thresholds fixed while changing the correlation gives a precise way to compare joint exceedance risk. This is aligned with the SOA Exam P (Probability) syllabus.

Definition: Let $(Z_1,Z_2)$ be a standard bivariate normal pair with correlation $\rho$, where $-1<\rho<1$. Each marginal variable has a standard normal distribution, and the pair has the bivariate normal density $$ \phi_2(x,y;\rho)=\frac{1}{2\pi\sqrt{1-\rho^2}}\exp\left(-\frac{x^2-2\rho xy+y^2}{2(1-\rho^2)}\right). $$ Write $\Phi$ for the standard normal cumulative distribution function and $\Phi_2(a,b;\rho)=P(Z_1\leq a,Z_2\leq b)$ for the bivariate standard normal CDF with correlation $\rho$.

For fixed thresholds $a$ and $b$, the event of interest is $Z_1>a$ and $Z_2>b$. Since each marginal is continuous, replacing either strict inequality by a non-strict one does not change its probability. Standardize first if the original variables are not standard normal: if $X$ and $Y$ have means $\mu_X,\mu_Y$, standard deviations $\sigma_X,\sigma_Y>0$, and correlation $\rho$, then the standardized thresholds are $a=(c-\mu_X)/\sigma_X$ and $b=(d-\mu_Y)/\sigma_Y$ for exceedance levels $c,d$.

Writing the Joint Upper Tail

The complement of the joint upper-tail event is the union of $Z_1\leq a$ and $Z_2\leq b$. Inclusion-exclusion therefore gives the general formula.

Theorem: For a standard bivariate normal pair with $-1<\rho<1$, $$ P(Z_1>a,Z_2>b)=1-\Phi(a)-\Phi(b)+\Phi_2(a,b;\rho). $$ At $\rho=0$, the variables are independent, so this reduces to $$ P(Z_1>a,Z_2>b)=[1-\Phi(a)][1-\Phi(b)]. $$

The formula is useful whether $\Phi_2$ is obtained from a calculator, software, or a table. A common computational error is to use $\Phi_2(a,b;\rho)$ itself as the probability of two exceedances. That CDF describes two values below their thresholds; the complements and inclusion-exclusion terms are essential.

Worked Example: Comparing Two Correlations at Zero Thresholds

Let $(Z_1,Z_2)$ be standard bivariate normal. Find $P(Z_1>0,Z_2>0)$ for $\rho=0.6$ and $\rho=-0.6$, and compare the results.

Method. For zero thresholds, the standard quadrant formula is $$ P(Z_1>0,Z_2>0)=\frac14+\frac{\arcsin(\rho)}{2\pi}. $$ Substitute each correlation. The formula incorporates dependence; multiplying the two marginal probabilities would only be justified at $\rho=0$.

For $\rho=0.6$, $\arcsin(0.6)=0.6435011$, so $$ \frac14+\frac{0.6435011}{2\pi} =0.25+\frac{0.6435011}{6.2831853} =0.25+0.1024164 =0.3524164. $$ To four decimal places, the probability is $0.3524$.

For $\rho=-0.6$, $\arcsin(-0.6)=-0.6435011$, so $$ \frac14-\frac{0.6435011}{6.2831853} =0.25-0.1024164 =0.1475836. $$ To four decimal places, the probability is $0.1476$. The results are consistent: positive correlation makes simultaneous positive values more likely than independence, while negative correlation makes them less likely. As a check, the two probabilities sum to $0.5000$, as required by the symmetry that changes both signs.

Common mistake avoided: Multiplying $0.5$ by $0.5$ gives $0.25$, which is correct only when $\rho=0$. It is not the answer for either nonzero correlation.

How the Probability Changes with Correlation

For fixed finite thresholds, increasing $\rho$ increases the joint upper-tail probability. The result can be justified by differentiating the bivariate normal CDF with respect to its correlation parameter. The key identity is $$ \frac{\partial}{\partial \rho}\Phi_2(a,b;\rho)=\phi_2(a,b;\rho). $$ Here is why it holds. Direct differentiation of the displayed bivariate density gives the identity $\partial\phi_2(x,y;\rho)/\partial\rho=\partial^2\phi_2(x,y;\rho)/(\partial x\,\partial y)$. Integrating both sides over $x\leq a,\ y\leq b$ turns the right side into its boundary value $\phi_2(a,b;\rho)$; the terms at negative infinity vanish because the density and its derivatives tend to zero there. This yields the CDF identity.

\frac{\partial}{\partial \rho}P(Z_1>a,Z_2>b) =\frac{\partial}{\partial \rho}\left[1-\Phi(a)-\Phi(b)+\Phi_2(a,b;\rho)\right] =\phi_2(a,b;\rho)>0.

The marginal terms do not depend on $\rho$, and the bivariate density is positive for $-1<\rho<1$. Thus the derivative is positive for every finite pair of thresholds. This comparison holds with the standardized thresholds fixed; if the thresholds change as well, this derivative alone does not compare the resulting probabilities.

Key takeaway: With both standardized thresholds held fixed, the joint upper-tail probability is strictly increasing in the correlation for $-1<\rho<1$. The marginal exceedance probabilities stay the same; only the dependence between the variables changes.

Independence and Limiting Cases

At zero correlation, a bivariate normal pair is independent. The joint upper-tail probability is then the product of its two marginal upper tails. At nonzero correlation, that product is generally not correct. For instance, even if one marginal exceedance is unlikely, positive dependence can increase the chance of observing both exceedances together.

The endpoint descriptions below are limits as $\rho$ approaches $1$ or $-1$. At these endpoints the bivariate normal distribution is degenerate, so the density formula for $-1<\rho<1$ does not apply directly. In the limits, the pair behaves like $(Z,Z)$ for positive perfect dependence and $(Z,-Z)$ for negative perfect dependence, with $Z$ standard normal.

As $\rho\to 1$, both exceedances require $Z>\max(a,b)$, so the limiting probability is $1-\Phi(\max(a,b))$.
As $\rho\to -1$, the conditions become $Z>a$ and $Z<-b$. Their intersection is the interval $a<Z<-b$, so the limiting probability is $\max\{\Phi(-b)-\Phi(a),0\}$.

The maximum with zero is important. If $a\geq -b$, the interval is empty, and the limiting probability is zero. If $a<-b$, it has positive length and its probability is the difference of the endpoint CDF values.

Worked Example: Fixed Thresholds and Three Dependence Regimes

A hypothetical monitoring system records two standardized measurements $Z_1,Z_2$ with a bivariate normal distribution. Compare $P(Z_1>0.5,Z_2>-0.5)$ at independence and in the limits as $\rho\to1$ and $\rho\to-1$.

At independence. The marginal exceedance probabilities are $$ P(Z_1>0.5)=1-\Phi(0.5)=0.3085375, $$ and $$ P(Z_2>-0.5)=1-\Phi(-0.5)=\Phi(0.5)=0.6914625. $$ Independence gives the product $$ 0.3085375(0.6914625)=0.2133421, $$ which is $0.2133$ to four decimal places.

As $\rho\to1$. The variables approach the same standard normal value. Both inequalities then require $Z>0.5$, since $0.5$ is the larger threshold. The limit is $1-\Phi(0.5)=0.3085375$, or $0.3085$ to four decimal places.

As $\rho\to-1$. The variables approach $Z$ and $-Z$. The two inequalities require $Z>0.5$ and $-Z>-0.5$, the latter equivalent to $Z<0.5$. These cannot both hold. Equivalently, the interval $0.5<Z<0.5$ is empty, so the limiting probability is $0.0000$.

The three results are consistent with the dependence comparison: the independent value lies between the limiting values. The negative-correlation limit is zero here because the thresholds leave no interval in which both conditions can hold.

Common Exam Traps

Using a product at nonzero correlation. The product of marginal tails applies at $\rho=0$, not in general.
Confusing lower and upper tails. $\Phi_2(a,b;\rho)$ is a lower-tail probability. Use inclusion-exclusion to obtain the joint upper tail.
Forgetting to standardize. For nonstandard normal variables, convert each original threshold using its own mean and standard deviation. The correlation is unchanged by this positive standardization.
Ignoring the interval at negative perfect dependence. The limiting probability is zero if the conditions $Z>a$ and $Z<-b$ are incompatible; it is not automatically a marginal tail probability.
Dropping a scale factor. In the zero-threshold quadrant formula, the arcsine term is divided by $2\pi$. Adding $\arcsin(\rho)$ directly to $1/4$ is incorrect.

Standardize the thresholds.
Convert each normal measurement to a standard normal variable and record its standardized exceedance level.

Identify the dependence model.
Write the correlation $\rho$ and check whether independence, a bivariate CDF, a quadrant formula, or a limiting case applies.

Calculate and interpret.
Use the correct upper-tail expression, retain sufficient precision during calculation, and check whether the result is consistent with the sign and size of the correlation.

Check Your Understanding

In each problem, keep the stated thresholds fixed when comparing correlations. Round probabilities to four decimal places.

For a standard bivariate normal pair with $\rho=0.4$, calculate $P(Z_1>0,Z_2>0)$. (Answer: $0.3155$)
For a standard bivariate normal pair with $\rho=-0.6$, calculate $P(Z_1>0,Z_2>0)$. (Answer: $0.1476$)
Suppose $Z_1,Z_2$ are independent standard normal variables. Calculate $P(Z_1>0.5,Z_2>1)$. (Answer: $0.0490$)
For fixed thresholds $a=0.5$ and $b=-0.5$, find the limits of $P(Z_1>a,Z_2>b)$ as $\rho\to1$ and as $\rho\to-1$. (Answer: as $\rho\to1$, $0.3085$; as $\rho\to-1$, $0.0000$)
For fixed thresholds $a=-0.5$ and $b=0.25$, find the limits of the joint upper-tail probability as $\rho\to1$ and as $\rho\to-1$. (Answer: as $\rho\to1$, $0.4013$; as $\rho\to-1$, $0.0928$)

What You'll Learn