Positive Covariance From a Shared Additive Component

Why a Shared Component Matters

Two measurements can move together because both contain the same random contribution. For example, two hypothetical project costs might each include a common randomly varying site-preparation expense, while their remaining costs differ. When that common expense is unusually high, both totals tend to be higher; when it is low, both tend to be lower. Covariance provides a numerical measure of this joint variation.

The key is to separate the common component from the parts that are not shared. A shared component contributes its variance to covariance when it enters both variables with positive coefficients. This contribution does not, by itself, prove that the total covariance is positive: dependence between the remaining parts can add to it or offset it.

Definition: Suppose $X=Z+U$ and $Y=Z+V$, where $Z$ is a random component included in both variables and $U,V$ are their respective residual components. By covariance bilinearity, $$ \operatorname{Cov}(X,Y) =\operatorname{Var}(Z)+\operatorname{Cov}(Z,V)+\operatorname{Cov}(U,Z)+\operatorname{Cov}(U,V). $$ If $Z$ is independent of the pair $(U,V)$, the two cross-covariances with $Z$ are zero, and $$ \operatorname{Cov}(X,Y)=\operatorname{Var}(Z)+\operatorname{Cov}(U,V). $$ If $U$ and $V$ are also independent, then $\operatorname{Cov}(U,V)=0$, so $\operatorname{Cov}(X,Y)=\operatorname{Var}(Z)$.

The independence conditions matter. Independence of $Z$ from each residual separately is enough to make the two cross-covariances zero, but the residuals $U$ and $V$ need not be independent of each other. Their covariance remains in the formula unless it is known to be zero.

Deriving the Shared-Component Formula

The general covariance identity follows by expanding around means. Write $\mu_Z=E[Z]$, $\mu_U=E[U]$, and $\mu_V=E[V]$. Then $X-E[X]=(Z-\mu_Z)+(U-\mu_U)$, and $Y-E[Y]=(Z-\mu_Z)+(V-\mu_V)$. Multiply these expressions and take expectations.

\begin{aligned} \operatorname{Cov}(X,Y) &=E[(X-E[X])(Y-E[Y])]\\ &=E[(Z-\mu_Z)^2] +E[(Z-\mu_Z)(V-\mu_V)]\\ &\quad +E[(U-\mu_U)(Z-\mu_Z)] +E[(U-\mu_U)(V-\mu_V)]\\ &=\operatorname{Var}(Z)+\operatorname{Cov}(Z,V) +\operatorname{Cov}(U,Z)+\operatorname{Cov}(U,V). \end{aligned}

This establishes why the common component contributes a variance: its centered deviation appears in both variables, so its product with itself is $(Z-E[Z])^2$, whose expectation is $\operatorname{Var}(Z)$. If $Z$ is independent of the residual pair, the cross-products have expectation zero. If the residuals are independent as well, their cross-product also has expectation zero.

A useful extension allows different positive or negative loadings on the shared component. If $X=aZ+U$ and $Y=bZ+V$, then $\operatorname{Cov}(X,Y)=ab\operatorname{Var}(Z)+a\operatorname{Cov}(Z,V)+b\operatorname{Cov}(U,Z)+\operatorname{Cov}(U,V)$. When $Z$ is independent of $(U,V)$, this reduces to $ab\operatorname{Var}(Z)+\operatorname{Cov}(U,V)$. A positive shared contribution requires $ab>0$ and a nonconstant $Z$. If either loading is zero, or $Z$ is constant, that contribution is zero; if the loadings have opposite signs, it is negative.

Worked Examples

Worked Example: Two Totals With a Shared Bernoulli Component

A hypothetical pair of operating totals is modeled as $X=Z+U$ and $Y=Z+V$. The shared component $Z$ is Bernoulli with $P(Z=1)=0.4$ and $P(Z=0)=0.6$. The residuals $U$ and $V$ are mutually independent and independent of $Z$, with $E[U]=3$, $E[V]=5$, $\operatorname{Var}(U)=2$, and $\operatorname{Var}(V)=7$. Find $\operatorname{Cov}(X,Y)$, and check it using expected products.

Method and assumptions. The stated mutual independence means every covariance between distinct components is zero. For a Bernoulli variable with success probability $p$, its variance is $p(1-p)$. Thus $\operatorname{Var}(Z)=0.4(0.6)=0.24$.

By the shared-component formula, $\operatorname{Cov}(X,Y)=\operatorname{Var}(Z)=0.24$. The positive covariance comes from the nonconstant component $Z$, not from the residual means or residual variances.

Check using the product-moment identity. First, $E[X]=E[Z]+E[U]=0.4+3=3.4$, and $E[Y]=E[Z]+E[V]=0.4+5=5.4$. Next, expand $XY=(Z+U)(Z+V)=Z^2+ZV+UZ+UV$. Since $Z^2=Z$ for a zero-one variable, $E[Z^2]=0.4$. Independence gives $E[ZV]=E[Z]E[V]=0.4(5)=2$, $E[UZ]=3(0.4)=1.2$, and $E[UV]=3(5)=15$. Therefore $E[XY]=0.4+2+1.2+15=18.6$, and $\operatorname{Cov}(X,Y)=E[XY]-E[X]E[Y]=18.6-(3.4)(5.4)=18.6-18.36=0.24$. Both methods agree.

Common mistake avoided: Do not treat the residual variances as terms in the covariance. They affect $\operatorname{Var}(X)$ and $\operatorname{Var}(Y)$, but do not enter $\operatorname{Cov}(X,Y)$ here because the residuals are independent of one another and of $Z$.

Worked Example: Different Loadings on the Shared Component

For two hypothetical claim-count indices, let $Z$ equal 0 or 6 with probability $1/2$ each. Define $X=2Z+U$ and $Y=3Z+V$. Assume $U$ and $V$ are independent of each other and of $Z$, with $E[U]=1$ and $E[V]=-2$. Find the covariance and verify it from $E[XY]$.

Method and assumptions. The loadings are $a=2$ and $b=3$, and all residual cross-covariances with the shared component are zero. The mean and second moment of $Z$ are $E[Z]=(0+6)/2=3$ and $E[Z^2]=(0^2+6^2)/2=18$. Hence $\operatorname{Var}(Z)=18-3^2=9$.

The scaled shared contribution is $ab\operatorname{Var}(Z)=2(3)(9)=54$. Since $\operatorname{Cov}(U,V)=0$, the covariance is $54$.

Check using expected products. The means are $E[X]=2(3)+1=7$ and $E[Y]=3(3)-2=7$. Expand $XY=6Z^2+2ZV+3ZU+UV$. Independence gives $E[2ZV]=2(3)(-2)=-12$, $E[3ZU]=3(3)(1)=9$, and $E[UV]=(1)(-2)=-2$. Also $E[6Z^2]=6(18)=108$. Thus $E[XY]=108-12+9-2=103$, so $\operatorname{Cov}(X,Y)=103-(7)(7)=103-49=54$. The direct check confirms the result.

Common mistake avoided: The shared contribution is $ab\operatorname{Var}(Z)$, not $ab(E[Z])^2$. Here $2(3)(E[Z])^2=54$ happens to give the same number, but that is coincidence: $E[Z]^2=9$ equals $\operatorname{Var}(Z)$ for this particular two-point distribution. The formula requires variance. For example, adding a constant to every possible value of $Z$ changes $E[Z]$ but does not change $\operatorname{Var}(Z)$ or the covariance contribution.

Worked Example: A Positive Shared Contribution With Dependent Residuals

Let $Z$ equal 0 or 4 with probability $1/2$ each, and let $W$ equal $-1$ or 1 with probability $1/2$ each, independently of $Z$. Set $X=Z+W$ and $Y=Z-W$. Find the covariance and explain why the shared component does not determine the answer by itself.

Method and assumptions. Here the residuals are $U=W$ and $V=-W$, so they are not independent. First, $E[Z]=2$, $E[Z^2]=(0+16)/2=8$, and $\operatorname{Var}(Z)=8-2^2=4$. Also $E[W]=0$, $E[W^2]=1$, and $\operatorname{Cov}(W,-W)=-E[W^2]=-1$.

Since $Z$ is independent of the residual pair, the cross-covariances with $Z$ are zero. Therefore $\operatorname{Cov}(X,Y)=\operatorname{Var}(Z)+\operatorname{Cov}(W,-W)=4-1=3$, which is positive.

Direct check. The means are $E[X]=E[Y]=2$. Also $XY=(Z+W)(Z-W)=Z^2-W^2$, so $E[XY]=8-1=7$. Therefore $\operatorname{Cov}(X,Y)=7-(2)(2)=3$, as found above.

Common mistake avoided: It would be incorrect to discard $\operatorname{Cov}(U,V)$ merely because the variables contain a shared component. The residuals here are dependent and contribute $-1$. Their negative covariance reduces, but does not eliminate, the shared component's contribution of 4.

When Is the Covariance Positive?

For $X=Z+U$ and $Y=Z+V$, if $Z$ is independent of the residual pair, the sign is determined by $\operatorname{Var}(Z)+\operatorname{Cov}(U,V)$. A nonconstant shared component supplies a strictly positive term, but the total covariance is positive only when $\operatorname{Cov}(U,V)>-\operatorname{Var}(Z)$. If the residual covariance equals the negative of the shared variance, the terms cancel; if it is more negative, the total covariance is negative.

When $U,V,Z$ are mutually independent, the residual covariance is zero, so any nonconstant $Z$ guarantees positive covariance in the equal-loading model. More generally, for independent residuals and $X=aZ+U,\;Y=bZ+V$, the covariance is $ab\operatorname{Var}(Z)$. Same-sign nonzero loadings yield a positive result; opposite-sign loadings yield a negative result.

Separate shared and residual terms.
Write each variable in a form such as $X=aZ+U$ and $Y=bZ+V$, identifying the shared random component explicitly.

Record the dependence assumptions.
Determine which component pairs are independent or have known covariance. Do not infer that $U$ and $V$ are independent just because both include $Z$.

Apply the expanded formula.
Use the appropriate variance and covariance terms, retaining any residual covariance or nonzero cross-covariance that the assumptions do not remove.

Check the sign and, if useful, verify directly.
Compare $E[XY]$ with $E[X]E[Y]$. A disagreement usually signals an omitted term or an arithmetic error.

This decomposition is useful in reliability, insurance, and operational models whenever two quantities respond to the same random environment or common shock. It also prevents an overclaim: positive association is not automatic merely because two variables share a term. The loadings and the dependence among residual terms must be considered.

Key takeaway: A common additive component $Z$ contributes $\operatorname{Var}(Z)$ to the covariance when it enters both variables once and is independent of the residual pair. With scaled loadings, the contribution is $ab\operatorname{Var}(Z)$. Residual covariance and cross-covariances must still be included whenever the assumptions do not make them zero.

Check Your Understanding

For each problem, identify the shared contribution and account for any residual covariance.

$Z$ takes the values 0, 2, and 4 with equal probabilities. Let $X=Z+U$ and $Y=Z+V$, where $U,V,Z$ are mutually independent. Find $\operatorname{Cov}(X,Y)$. (Answer: $8/3$)
$Z$ is Bernoulli with $P(Z=1)=0.25$. Let $X=4Z+U$ and $Y=2Z+V$, where $U,V,Z$ are mutually independent. Find $\operatorname{Cov}(X,Y)$. (Answer: $1.5$)
$Z$ and the residual pair $(U,V)$ are independent, with $\operatorname{Var}(Z)=5$ and $\operatorname{Cov}(U,V)=-2$. For $X=Z+U,\;Y=Z+V$, find $\operatorname{Cov}(X,Y)$. (Answer: $3$)
$Z$ takes the values 1 and 5 with equal probabilities. Let $X=Z+U$, $Y=2Z+V$, and assume $U,V,Z$ are mutually independent. Find $\operatorname{Cov}(X,Y)$. (Answer: $8$)
$Z$ takes the values 0 and 6 with equal probabilities. Let $X=2Z+U$ and $Y=3Z+V$, where $U,V,Z$ are mutually independent. Find $\operatorname{Cov}(X,Y)$. (Answer: $54$)

What You'll Learn