Vulnerability Uncertainty and Damage Variability

Why a Vulnerability Curve Does Not Tell the Whole Story

A vulnerability curve gives an expected damage ratio for a specified hazard intensity. For example, a curve might assign a mean damage ratio of 0.30 at a particular intensity. That value describes an average under the model; it does not say that every asset will sustain exactly 30% damage, nor does it guarantee that the mean itself is known precisely.

These are two distinct sources of uncertainty. Damage variability is the variation in actual damage around the curve’s mean, conditional on the mean being specified. Vulnerability-curve uncertainty is uncertainty about what that mean should be at the specified intensity, often because the curve’s parameters or supporting data are uncertain. Keeping the two separate matters when estimating the spread of modeled loss, and when representing the dependence of damage across assets.

Definition: Fix a hazard intensity $h$ and let $R$ be an asset’s damage ratio, between 0 and 1. Let $M$ be the uncertain mean damage ratio at that intensity. The model specifies $E[R\mid M]=M$. The conditional variance $\operatorname{Var}(R\mid M)$ represents damage variability around the specified mean; $\operatorname{Var}(M)$ represents uncertainty in the mean itself. If the asset’s value is $V$, its modeled damage loss is $X=VR$.

In a fitted vulnerability model, one can represent the mean as $M=m(h;\theta)$, where $m$ is the curve and $\theta$ denotes uncertain curve parameters. For a fixed $h$, uncertainty in $\theta$ induces uncertainty in $M$. The calculations below use simple discrete distributions so each contribution can be seen directly.

Separate the Two Contributions to Variance

The model’s expected damage ratio is $E[R]=E[M]$. Its variance has two components: average damage variability conditional on the mean, and variability in the mean itself.

\operatorname{Var}(R) = E\!\left[\operatorname{Var}(R\mid M)\right] + \operatorname{Var}\!\left(E[R\mid M]\right) = E\!\left[\operatorname{Var}(R\mid M)\right] + \operatorname{Var}(M).

This is the law of total variance. For dollar damage $X=VR$, where $V$ is fixed, the corresponding results are $E[X]=V E[M]$ and $\operatorname{Var}(X)=V^2\operatorname{Var}(R)$. Variance is multiplied by $V^2$, not by $V$.

Condition on the mean.
For each possible value of $M$, calculate the conditional mean $E[R\mid M]$ and conditional variance $\operatorname{Var}(R\mid M)$.

Average the conditional variance.
Weight each conditional variance by the probability of its value of $M$. This gives $E[\operatorname{Var}(R\mid M)]$.

Calculate uncertainty in the mean.
Find $\operatorname{Var}(M)$ across the possible curve means.

Add the components and convert to loss units.
Add the two ratio-variance components. Multiply the resulting variance by $V^2$ to obtain dollar-loss variance.

The identity follows by writing $R-E[R]=(R-E[R\mid M])+(E[R\mid M]-E[R])$ and squaring. The cross-product has expectation zero: conditional on $M$, $E[R-E[R\mid M]\mid M]=0$, while $E[R\mid M]-E[R]$ is fixed. The remaining squared terms have expectations $E[\operatorname{Var}(R\mid M)]$ and $\operatorname{Var}(E[R\mid M])$, respectively. Since $E[R\mid M]=M$, the second term is $\operatorname{Var}(M)$.

Worked Calculations

Worked Example: One Asset With an Uncertain Curve Mean

A hypothetical building has value $V=\$100{,}000$. For a fixed hazard intensity, its uncertain mean damage ratio $M$ equals 0.20 or 0.40, each with probability 0.50. Conditional on $M$, the damage ratio $R$ equals $M-0.10$ or $M+0.10$, each with probability 0.50. Calculate expected dollar damage, its variance, and its standard deviation.

Method. Apply total expectation and total variance, first in ratio units and then in dollars. This discrete conditional model keeps the damage ratios between 0 and 1.

The mean of $M$ is

$$ E[M]=0.50(0.20)+0.50(0.40)=0.30. $$

Thus $E[R]=0.30$, and expected dollar damage is

E[X]=100{,}000(0.30)=\$30{,}000.

At either value of $M$, $R$ is equally likely to be 0.10 below or 0.10 above that mean. Therefore $\operatorname{Var}(R\mid M)=0.10^2=0.01$ for both values, so $E[\operatorname{Var}(R\mid M)]=0.01$. The variance of the uncertain mean is

\operatorname{Var}(M) =0.50(0.20-0.30)^2+0.50(0.40-0.30)^2 =0.01.

The total ratio variance is $0.01+0.01=0.02$. In dollar units,

\operatorname{Var}(X)=100{,}000^2(0.02)=200{,}000{,}000\text{ dollars}^2, \qquad \operatorname{SD}(X)=\sqrt{200{,}000{,}000}=\$14{,}142.14.

The standard deviation is rounded to the nearest cent. As a check, the unconditional ratios are 0.10, 0.30, 0.30, and 0.50, each with probability 0.25. Their mean is 0.30 and their variance is 0.02, as calculated.

Worked Example: Damage Variability With a Known Mean

A hypothetical commercial asset has value $V=\$200{,}000$, and its vulnerability-curve mean at a specified intensity is known to be $M=0.30$. Conditional on this mean, the damage ratio $R$ is 0.10 with probability 0.25, 0.30 with probability 0.50, and 0.50 with probability 0.25. Find expected dollar damage, dollar-loss variance, and standard deviation.

Method. Because the mean is known, there is no variance in $M$; calculate only the conditional damage variability. The common mistake this avoids is treating the curve mean as the realized damage ratio.

First verify the conditional mean:

E[R\mid M=0.30] =0.25(0.10)+0.50(0.30)+0.25(0.50) =0.025+0.150+0.125=0.30.

Expected dollar damage is therefore

E[X]=200{,}000(0.30)=\$60{,}000.

The conditional variance is the probability-weighted squared deviation from 0.30:

\operatorname{Var}(R\mid M=0.30) =0.25(0.10-0.30)^2+0.50(0.30-0.30)^2+0.25(0.50-0.30)^2 =0.25(0.04)+0+0.25(0.04)=0.02.

Since $M$ is fixed, $\operatorname{Var}(M)=0$, so the total ratio variance is 0.02. Converting to dollars gives

\operatorname{Var}(X)=200{,}000^2(0.02)=800{,}000{,}000\text{ dollars}^2, \qquad \operatorname{SD}(X)=\sqrt{800{,}000{,}000}=\$28{,}284.27.

The standard deviation is rounded to the nearest cent. The result is consistent with the distribution: ratios vary around 0.30 even though the curve mean is known exactly.

Worked Example: Shared Curve Uncertainty Across Two Assets

Two hypothetical buildings each have value $\$100{,}000$ and experience the same fixed hazard intensity. Their common uncertain mean damage ratio $M$ is 0.20 or 0.40, each with probability 0.50. Conditional on $M$, their damage ratios $R_1$ and $R_2$ are independent; each equals $M-0.10$ or $M+0.10$ with probability 0.50 each. Find the expected portfolio damage and its variance.

Method. Condition on the shared mean first. Conditional independence applies to the residual damage variability, not to the common curve mean. This distinction avoids incorrectly treating curve uncertainty as independent for each building.

Each ratio has $E[R_j\mid M]=M$ and $\operatorname{Var}(R_j\mid M)=0.01$. Let $X=100{,}000(R_1+R_2)$ be the portfolio damage. Its expectation is

E[X]=100{,}000\{E[R_1]+E[R_2]\} =100{,}000(0.30+0.30)=\$60{,}000.

Conditional on $M$, the ratios are independent, so their variances add:

E[\operatorname{Var}(X\mid M)] =100{,}000^2\{0.01+0.01\} =200{,}000{,}000\text{ dollars}^2.

The conditional mean of portfolio damage is $E[X\mid M]=200{,}000M$. Since $\operatorname{Var}(M)=0.01$,

\operatorname{Var}(E[X\mid M]) =200{,}000^2(0.01) =400{,}000{,}000\text{ dollars}^2.

The total portfolio variance is the sum:

\operatorname{Var}(X) =200{,}000{,}000+400{,}000{,}000 =600{,}000{,}000\text{ dollars}^2, \qquad \operatorname{SD}(X)=\$24{,}494.90.

The standard deviation is rounded to the nearest cent. As a check, each building’s ratio has variance $0.01+0.01=0.02$, and the two ratios have covariance $\operatorname{Var}(M)=0.01$. Thus the sum’s ratio variance is $0.02+0.02+2(0.01)=0.06$; multiplying by $100{,}000^2$ also gives $600{,}000{,}000$ dollars squared.

Interpreting the Results and Avoiding a Trap

The curve mean and damage variability answer different questions. Curve-mean uncertainty describes uncertainty about the expected damage at the specified intensity. Damage variability describes the range of possible realized damage even if that expected value were known. A model may include one source, both, or neither; the assumptions should state which sources are represented.

A frequent error is to report only the first conditional variance and omit uncertainty in the curve mean. In the first worked example, using $0.01$ alone would give the wrong dollar-loss variance:

\text{Wrong: }100{,}000^2(0.01)=100{,}000{,}000\text{ dollars}^2. \qquad \text{Correct: }100{,}000^2(0.01+0.01)=200{,}000{,}000\text{ dollars}^2.

The reverse error is also possible: adding an uncertainty-in-mean term when the curve mean is specified as fixed. In that case $\operatorname{Var}(M)=0$, so only conditional damage variability contributes. In either case, use the model’s stated assumptions rather than automatically adding two nonzero terms.

For multiple assets exposed to the same hazard, a shared uncertain curve mean can also create dependence. Even if conditional damage variability is independent across assets given $M$, the common $M$ affects their conditional expected damage together. Treating the uncertain mean as a separate independent draw for each asset would remove that shared component and can understate portfolio variance. This treatment concerns modeled damage at a fixed intensity; it does not by itself specify event frequency, policy terms, or insured loss.

Key takeaway: At a fixed hazard intensity, write the conditional mean and conditional variance of damage explicitly. Then use $\operatorname{Var}(R)=E[\operatorname{Var}(R\mid M)]+\operatorname{Var}(M)$. For fixed asset value $V$, multiply ratio variance by $V^2$; for a shared uncertain mean across assets, account for the resulting dependence.

Check Your Understanding

For each problem, distinguish the conditional damage variability from uncertainty in the vulnerability-curve mean. Monetary standard deviations are rounded to the nearest cent.

A hypothetical asset worth $\$80{,}000$ has a known mean damage ratio of 0.25. Conditional on that mean, its ratio is 0.15 or 0.35 with equal probability. Find expected damage and dollar-loss variance. (Answer: expected damage $\$20{,}000$; variance $64{,}000{,}000$ dollars squared.)
A hypothetical asset worth $\$150{,}000$ has uncertain mean ratio $M=0.20$ or $0.40$, each with probability 0.50. Given $M$, its ratio is $M-0.05$ or $M+0.05$, equally likely. Find expected damage and dollar-loss variance. (Answer: expected damage $\$45{,}000$; variance $281{,}250{,}000$ dollars squared.)
Two hypothetical assets, each worth $\$50{,}000$, share an uncertain mean ratio $M=0.10$ or $0.30$, equally likely. Given $M$, their ratios are independent and each is $M-0.05$ or $M+0.05$, equally likely. Find expected portfolio damage and variance. (Answer: expected damage $\$20{,}000$; variance $112{,}500{,}000$ dollars squared.)
A hypothetical asset worth $\$300{,}000$ has a fixed mean ratio of 0.40. Its ratio is 0.20, 0.40, or 0.60 with probabilities 0.25, 0.50, and 0.25. Find expected damage and dollar-loss variance. (Answer: expected damage $\$120{,}000$; variance $1{,}800{,}000{,}000$ dollars squared.)
Two assets share uncertain mean $M$. Given $M$, their damage ratios are independent and both have conditional mean $M$. In one sentence, identify the source of their unconditional dependence. (Answer: the shared random mean $M$ induces dependence because both conditional expected damage ratios move together.)

What You'll Learn