Why Covariance Needs Units
Covariance summarizes how two variables’ deviations from their means behave together. Its sign can tell us whether their deviations tend to contribute products of the same or opposite signs. Its numerical value, however, is not a unit-free score: it carries the units of both variables multiplied together. This matters whenever variables represent quantities such as dollars and hours, or when the same data are reported in different measurement units.
For example, if \(X\) is measured in dollars and \(Y\) in days, then \(\operatorname{Cov}(X,Y)\) is measured in dollar-days. If a cost is instead recorded in hundreds of dollars, the numerical covariance changes even though the underlying physical cost has not changed. A covariance’s magnitude therefore needs its units and measurement scales to be interpreted.
The definition makes the units visible: each centered value \(X-E[X]\) has the same units as \(X\), and each centered value \(Y-E[Y]\) has the same units as \(Y\). Their product, and hence its expected value, has the product of those units. If both variables are measured in the same unit, that unit is squared. For instance, a variance of time measured in minutes is in minutes squared, not minutes.
Worked Example: Covariance From a Joint Probability Model
A hypothetical claims team models claim size \(X\), in hundreds of dollars, and processing time \(Y\), in hours, using the following discrete joint distribution. Find the covariance, state its units, and express it in dollar-hours. Assume the listed outcomes and probabilities describe the full distribution; no interest or mortality assumption is involved.
| \(X\) | \(Y\) | Probability |
|---|---|---|
| 1 | 1 | 0.4 |
| 1 | 3 | 0.1 |
| 3 | 1 | 0.1 |
| 3 | 3 | 0.4 |
Method. Find the two means, then average the products of the centered outcomes. The probabilities sum to \(0.4+0.1+0.1+0.4=1.0\), so the table is a valid distribution.
For \(X\), $$ E[X]=(1)(0.4)+(1)(0.1)+(3)(0.1)+(3)(0.4)=0.4+0.1+0.3+1.2=2. $$ For \(Y\), $$ E[Y]=(1)(0.4)+(3)(0.1)+(1)(0.1)+(3)(0.4)=0.4+0.3+0.1+1.2=2. $$
The centered products for the four outcomes are \((1-2)(1-2)=1\), \((1-2)(3-2)=-1\), \((3-2)(1-2)=-1\), and \((3-2)(3-2)=1\). Weighting each by its probability gives $$ \operatorname{Cov}(X,Y)=(1)(0.4)+(-1)(0.1)+(-1)(0.1)+(1)(0.4)=0.6. $$
The result is \(0.6\) hundred-dollar-hours. Since one hundred dollars is \(100\) dollars, the covariance is \(0.6(100)=60\) dollar-hours. As a check, \(E[XY]=0.4(1)+0.1(3)+0.1(3)+0.4(9)=4.6\), while \(E[X]E[Y]=2(2)=4\); their difference is \(4.6-4=0.6\) hundred-dollar-hours. The product-moment check agrees. A common mistake is to report \(0.6\) as though it were already in dollar-hours; that omits the factor of \(100\) in the cost unit.
Why the Product-Moment Formula Holds
The identity \(\operatorname{Cov}(X,Y)=E[XY]-E[X]E[Y]\) is useful for calculations, but it also makes the unit behavior clear: \(XY\) and \(E[X]E[Y]\) both have the product of the variables’ units. The identity follows by expanding the definition and using linearity of expectation.
\(\operatorname{Cov}(X,Y)=E[(X-E[X])(Y-E[Y])]\).
\((X-E[X])(Y-E[Y])=XY-XE[Y]-YE[X]+E[X]E[Y]\).
Since \(E[X]\) and \(E[Y]\) are constants, linearity gives \(E[XY]-E[Y]E[X]-E[X]E[Y]+E[X]E[Y]\).
The terms combine to give \(\operatorname{Cov}(X,Y)=E[XY]-E[X]E[Y]\).
This derivation assumes the expectations involved are finite; finite variances for \(X\) and \(Y\) ensure the covariance is defined. It also shows why one should not subtract quantities with different units: both terms on the right must have the same product units.
How Unit Changes Affect Covariance
A change of measurement unit multiplies the numerical value of a variable by a conversion factor. Thus, converting one variable changes covariance by that factor; converting both variables changes it by the product of their factors. Adding a constant to a variable, by contrast, does not change covariance, because centering removes the constant.
To justify the result, first center the transformed variables: \(aX+b-E[aX+b]=a(X-E[X])\), and similarly \(cY+d-E[cY+d]=c(Y-E[Y])\). Their product is \(ac(X-E[X])(Y-E[Y])\). Taking expectations gives the stated formula. For a variance, setting both variables equal to \(X\) gives \(\operatorname{Var}(aX+b)=a^2\operatorname{Var}(X)\).
Worked Example: Converting Liters and Minutes
In a hypothetical warehouse model, package volume \(X\) is measured in liters and handling time \(Y\) in minutes. The joint distribution has four equally specified outcomes with probabilities shown below. Find the covariance in liter-minutes, then convert the result to milliliter-hours. This is a discrete probability model; no interest or mortality assumption is involved.
| \(X\) (liters) | \(Y\) (minutes) | Probability |
|---|---|---|
| 2 | 10 | 0.3 |
| 4 | 10 | 0.2 |
| 2 | 30 | 0.2 |
| 4 | 30 | 0.3 |
Method. Compute the means and centered-product expectation, then apply both unit-conversion factors. The probabilities sum to \(1.0\).
The means are $$ E[X]=2(0.3)+4(0.2)+2(0.2)+4(0.3)=0.6+0.8+0.4+1.2=3 $$ and $$ E[Y]=10(0.3)+10(0.2)+30(0.2)+30(0.3)=3+2+6+9=20. $$ The four centered products are \((2-3)(10-20)=10\), \((4-3)(10-20)=-10\), \((2-3)(30-20)=-10\), and \((4-3)(30-20)=10\). Therefore, $$ \operatorname{Cov}(X,Y)=10(0.3)+(-10)(0.2)+(-10)(0.2)+10(0.3)=3-2-2+3=2 $$ liter-minutes.
One liter is \(1{,}000\) milliliters, and one minute is \(1/60\) hour. The combined conversion factor is \(1{,}000(1/60)=50/3\), so $$ \operatorname{Cov}(X_{\text{mL}},Y_{\text{hours}})=2(50/3)=100/3 $$ milliliter-hours, or \(33.33\) milliliter-hours to two decimal places. The unit factors multiply; dividing by \(60\) for the time conversion does not mean dividing the entire covariance by \(60\) and forgetting the volume conversion. The result is checked by the conversion rule: \(2\) liter-minutes corresponds to \(2(1{,}000)(1/60)=100/3\) milliliter-hours.
Why Magnitude Alone Is Not Association Strength
Covariance depends on the units and numerical scales of both variables. A large absolute value can arise because the variables are recorded in large units, because their values vary widely, or because their centered values tend to move together. The number alone does not separate these effects. Even for the same physical quantities, changing from dollars to hundreds of dollars changes the numerical covariance by a factor of \(100\).
A useful exam habit is to write the units beside every covariance as soon as it is calculated. Then ask whether the variables and measurement scales are the same before comparing two covariance values. A covariance in dollar-days cannot be compared numerically with one in thousands-of-dollars-minutes as if they were on a common scale. Covariance is not standardized, so there is no universal numerical benchmark that makes a value “large” or “small” across different variable scales.
Worked Example: A Shift Does Not Change Covariance, but a Scale Does
A hypothetical production monitor records two binary indicators, \(U\) and \(V\): each is either \(0\) or \(1\). Let their joint distribution be \(P(U=0,V=0)=0.4\), \(P(U=0,V=1)=0.1\), \(P(U=1,V=0)=0.1\), and \(P(U=1,V=1)=0.4\). A cost variable is \(C=500+100U\) dollars, and a delay variable is \(D=1+2V\) days. Find \(\operatorname{Cov}(C,D)\), and compare it with the covariance when cost is recorded in hundreds of dollars. This is a discrete model with the stated probabilities and transformations; no interest or mortality assumption is involved.
Method. First calculate the covariance of the indicators, then use the affine transformation rule. The common mistake is to treat the added constants as if they contributed to covariance.
The marginal means are \(E[U]=P(U=1)=0.1+0.4=0.5\) and \(E[V]=P(V=1)=0.1+0.4=0.5\). Also, \(E[UV]=P(U=1,V=1)=0.4\), so $$ \operatorname{Cov}(U,V)=E[UV]-E[U]E[V]=0.4-(0.5)(0.5)=0.15. $$ Since \(C=100U+500\) and \(D=2V+1\), the scale factors are \(100\) and \(2\); the shifts \(500\) and \(1\) have no effect. Thus, $$ \operatorname{Cov}(C,D)=(100)(2)(0.15)=30 $$ dollar-days.
Now express cost in hundreds of dollars: \(C/100=5+U\). Its covariance with \(D\) is $$ \operatorname{Cov}(C/100,D)=(1)(2)(0.15)=0.3 $$ hundred-dollar-days. The physical relationship described by these transformed quantities has not changed, but the numerical covariance is \(100\) times smaller because the cost unit is \(100\) times larger. The conversion check is \(30\) dollar-days divided by \(100\), which equals \(0.3\) hundred-dollar-days. Reporting \(30\) and \(0.3\) without their units would invite a misleading comparison.
Common Exam Traps
- Dropping the product units. If \(X\) is in dollars and \(Y\) is in hours, write dollar-hours, not dollars or hours.
- Giving variance the original unit. If \(X\) is measured in minutes, \(\operatorname{Var}(X)\) is measured in minutes squared.
- Converting only one variable. If both variables’ units change, multiply covariance by both conversion factors.
- Letting a shift affect covariance. Adding a fixed amount changes the mean but leaves centered values, and therefore covariance, unchanged.
- Interpreting a large number without its scale. First check the units and the scales of both variables; the magnitude alone is not a unit-free measure.
Check Your Understanding
For each problem, track the units as well as the numerical calculation.
- A covariance is \(-2.4\) kilometer-hours. Convert it to meter-minutes. (Answer: \(-144{,}000\) meter-minutes)
- A variable \(T\) measured in minutes has variance \(900\) minutes squared. What is its variance when \(T\) is expressed in hours? (Answer: \(0.25\) hours squared)
- \(\operatorname{Cov}(A,B)=7\) meter-kilograms. Let \(X=3A+4\) and \(Y=0.5B-9\), where the constants are numerical scale factors and shifts. Find \(\operatorname{Cov}(X,Y)\) and its units. (Answer: \(10.5\) meter-kilograms)
- \(\operatorname{Cov}(X,Y)=-12\) dollar-hours. Express \(X\) in thousands of dollars by using \(X/1{,}000\), and express \(Y\) in minutes by using \(60Y\). Find the transformed covariance. (Answer: \(-0.72\) thousand-dollar-minutes)
- Two measurements describe the same pair of physical variables. Their covariance is reported as \(45\) dollar-days and \(0.45\) hundred-dollar-days. Is one covariance evidence of stronger association? Explain briefly. (Answer: No; these are equivalent values in different cost units, since \(45\) dollars equals \(0.45\) hundred dollars.)