The Delta Method for Parameter Functions

Why Transform an Estimate?

Actuarial work often requires a quantity that is not itself a model parameter. A fitted loss model may be specified by a mean parameter, for example, while a calculation needs its reciprocal, its logarithm, or the variance implied by that parameter. An estimate of the parameter does not automatically tell us how uncertain the transformed estimate is. The delta method provides a first-order approximation by expressing a smooth transformation as a line near the estimated parameter.

The method is useful when an estimator is approximately normal for a large sample and the function of interest is differentiable. It does not make a nonlinear transformation exactly normal, nor does it guarantee a good approximation in a small sample or near a point where the function changes sharply.

Definition: Suppose $\hat{\theta}$ estimates a scalar parameter $\theta$, and for sample size $n$, $\sqrt{n}(\hat{\theta}-\theta)$ converges in distribution to a normal random variable with mean $0$ and variance $V(\theta)$. For a function $g$ differentiable at $\theta$, the first-order delta method gives $\sqrt{n}\{g(\hat{\theta})-g(\theta)\}$ an approximate normal distribution with mean $0$ and variance $[g'(\theta)]^2V(\theta)$. Equivalently, $\operatorname{Var}\{g(\hat{\theta})\}$ is approximately $[g'(\theta)]^2V(\theta)/n$.

In applications, the unknown $\theta$ in the variance formula is replaced by $\hat{\theta}$, and $V(\theta)$ is replaced by an appropriate estimate or model-based expression. The resulting standard error is an estimate of the standard deviation of $g(\hat{\theta})$, not the standard deviation of the original observations.

Derivation by Linearization

The central idea is to approximate the change in $g$ by the slope of its tangent line. If $g$ is differentiable at $\theta$, then for $\hat{\theta}$ near $\theta$,

g(\hat{\theta}) \approx g(\theta)+g'(\theta)(\hat{\theta}-\theta).

Subtracting $g(\theta)$ shows that the estimation error on the transformed scale is approximately the original estimation error multiplied by $g'(\theta)$. Multiplication by a constant multiplies variance by the square of that constant. Thus,

\operatorname{Var}\{g(\hat{\theta})\} \approx [g'(\theta)]^2\operatorname{Var}(\hat{\theta}) \approx [g'(\theta)]^2\frac{V(\theta)}{n}.

The same reasoning gives the asymptotic distribution. The Taylor expansion has a remainder that is small relative to $\hat{\theta}-\theta$ when $\hat{\theta}$ is consistent and $g$ is sufficiently smooth near $\theta$. Under the stated asymptotic normality, the leading term is therefore $g'(\theta)\sqrt{n}(\hat{\theta}-\theta)$. A normal limit with variance $V(\theta)$ becomes a normal limit with variance $[g'(\theta)]^2V(\theta)$.

Theorem: If $\hat{\theta}$ is consistent for $\theta$, $\sqrt{n}(\hat{\theta}-\theta)$ is asymptotically normal with variance $V(\theta)$, and $g$ is differentiable at $\theta$, then the first-order delta-method variance estimate is $\widehat{\operatorname{Var}}\{g(\hat{\theta})\}=[g'(\hat{\theta})]^2\widehat{\operatorname{Var}}(\hat{\theta})$. The corresponding estimated standard error is $|g'(\hat{\theta})|\widehat{\operatorname{SE}}(\hat{\theta})$.

The absolute value appears in the standard error because a standard deviation cannot be negative. The sign of $g'(\hat{\theta})$ still describes the direction of change: if the function is decreasing, larger parameter estimates produce smaller transformed estimates.

Applying the Approximation

A reliable calculation follows the same sequence each time: identify the parameter and its estimator, write the target as a function $g$, differentiate, and combine the squared derivative with the estimator variance. The following hypothetical examples use model assumptions stated explicitly; all dollar amounts refer to the stated hypothetical loss setting.

Worked Example: Transforming an Exponential Mean Into a Rate

A hypothetical portfolio has independent, identically distributed positive losses $X_1,\ldots,X_n$ with exponential density $f_X(x)=\theta^{-1}e^{-x/\theta}$ for $x>0$, where $\theta>0$ is the mean loss in dollars. The maximum likelihood estimator is the sample mean. For $n=100$, suppose the sample mean is $\hat{\theta}=\$2{,}000$. Estimate the reciprocal parameter $g(\theta)=1/\theta$, and its standard error.

Method and assumptions. For this model, $E[X]=\theta$ and $\operatorname{Var}(X)=\theta^2$. The sample mean is asymptotically normal, with $\operatorname{Var}(\hat{\theta})\approx\theta^2/n$. We use the plug-in estimate $\widehat{\operatorname{Var}}(\hat{\theta})=\hat{\theta}^2/n$.

Step 1: Estimate the target. The target is the exponential rate, expressed per dollar: $g(\hat{\theta})=1/2{,}000=0.0005$ per dollar.

Step 2: Differentiate. Since $g(\theta)=\theta^{-1}$, its derivative is $g'(\theta)=-\theta^{-2}$. At the estimate, $g'(\hat{\theta})=-1/2{,}000^2=-1/4{,}000{,}000$.

Step 3: Estimate the variance and standard error. The estimated variance of the mean is $2{,}000^2/100=40{,}000$ dollars squared. Therefore,

\widehat{\operatorname{Var}}\{g(\hat{\theta})\} =\left(\frac{1}{2{,}000^2}\right)^2(40{,}000) =2.5\times10^{-9}\ \text{dollars}^{-2}, \qquad \widehat{\operatorname{SE}}\{g(\hat{\theta})\} =\sqrt{2.5\times10^{-9}} =0.00005\ \text{per dollar}.

As an optional normal-approximation check, a 95% interval on the rate scale is $0.0005\pm1.96(0.00005)=[0.000402,0.000598]$ per dollar. The interval is positive, consistent with a positive rate parameter.

Check and common mistake. The estimated standard error is one tenth of the estimated rate, as expected from the relative standard error of a sample mean with $n=100$. A common error is to use $\operatorname{Var}(\hat{\theta})$ alone as the variance of the rate estimate. The derivative factor $[g'(\hat{\theta})]^2$ is essential. The reported variance has units of inverse dollars squared, while the standard error has units of inverse dollars.

Worked Example: Uncertainty in a Lognormal Median

Suppose hypothetical claim severities follow a lognormal model: the natural logarithm of each severity, measured numerically in dollars, is normally distributed with mean $\mu$ and known standard deviation $\sigma=0.4$. For $n=64$ independent claims, the sample mean of the log severities is $\hat{\mu}=\ln(1{,}000)$. Estimate the median severity and its delta-method standard error.

Method and assumptions. With $\sigma$ known, the sample mean of the log severities estimates $\mu$, and its variance is $\sigma^2/n$. The median of a lognormal distribution is $g(\mu)=e^\mu$, so the estimated median is $e^{\hat{\mu}}$.

Step 1: Estimate the median. Since $\hat{\mu}=\ln(1{,}000)$, $g(\hat{\mu})=e^{\ln(1{,}000)}=\$1{,}000$.

Step 2: Differentiate and estimate the variance. The derivative is $g'(\mu)=e^\mu$, hence $g'(\hat{\mu})=1{,}000$. The estimated variance of $\hat{\mu}$ is $0.4^2/64=0.16/64=0.0025$. Therefore the approximate variance of the estimated median is $1{,}000^2(0.0025)=2{,}500$ dollars squared.

Step 3: Take the square root. The standard error is $\sqrt{2{,}500}=\$50$. A 95% symmetric normal approximation on the median scale is $\$1{,}000\pm1.96(\$50)$, or approximately $[\$902,\$1{,}098]$, rounded to the nearest dollar.

Check and common mistake. The standard error is 5% of the estimated median, matching the relative standard deviation $0.4/\sqrt{64}=0.05$ of $\hat{\mu}$. Do not treat the median as $e^{\hat{\mu}}$ but leave its uncertainty on the log scale: the derivative $e^{\hat{\mu}}$ converts log-scale uncertainty into dollar-scale uncertainty. This interval is a first-order approximation, not an exact lognormal confidence interval.

Worked Example: Estimating Gamma Severity Variance

A hypothetical loss model has a gamma distribution with known shape $k=2$ and unknown scale $\theta>0$, so $E[X]=k\theta$ and $\operatorname{Var}(X)=k\theta^2$. For $n=50$ independent losses, suppose the sample mean is $\$600$. Estimate the model variance and the standard error of that estimate.

Method and assumptions. With known shape, the maximum likelihood estimator of the scale is $\hat{\theta}=\bar{X}/k$. The variance of one gamma loss is $k\theta^2$, so the large-sample variance of $\hat{\theta}$, obtained from the sample mean, is approximately $\theta^2/(nk)$.

Step 1: Estimate the scale and target. The scale estimate is $\hat{\theta}=600/2=\$300$. The target function is $g(\theta)=k\theta^2$, giving the estimated loss variance $g(\hat{\theta})=2(300^2)=\$180{,}000$ squared.

Step 2: Find the derivative and parameter-estimator variance. Here $g'(\theta)=2k\theta$, so $g'(\hat{\theta})=2(2)(300)=1{,}200$. The estimated variance of $\hat{\theta}$ is $300^2/(50\cdot2)=90{,}000/100=900$ dollars squared.

Step 3: Apply the delta method. The approximate variance of the estimated model variance is $1{,}200^2(900)=1{,}440{,}000(900)=1{,}296{,}000{,}000$ dollars to the fourth power. Its standard error is $\sqrt{1{,}296{,}000{,}000}=\$36{,}000$ squared.

Check and common mistake. The target is a variance, so its standard error has units of dollars squared. The number $900$ is the variance of the estimated scale; it is not the variance of the estimated loss variance. The derivative $2k\hat{\theta}$ is required to propagate uncertainty to the target.

Conditions, Interpretation, and Exam Traps

The first-order approximation is most useful when the estimator is concentrated near the true parameter and the function is smooth over that region. A large derivative means a small parameter-estimation error can create a substantial change in the target; a small derivative means the transformed estimate is locally less sensitive. Always evaluate the derivative at the estimate when calculating a plug-in standard error.

A differentiable function need not yield a good normal approximation in every finite sample. For example, a symmetric normal interval for a positive quantity may extend below zero, or the function may be strongly curved across the likely range of $\hat{\theta}$. Such warnings do not change the delta-method calculation; they indicate that the first-order approximation may be inadequate for the intended inference.

The first-order method also requires care if $g'(\theta)=0$. In that case, its variance formula gives zero, but the transformed estimate can still vary through the function's curvature. A second-order expansion may then be needed; the zero first-order result should not be interpreted as proof of zero uncertainty.

Specify the estimator's uncertainty.
Write down the approximate variance of $\hat{\theta}$, using the stated model and sample size.

Define the target function.
Express the requested quantity as $g(\theta)$ so the transformation is explicit.

Differentiate and substitute.
Calculate $g'(\hat{\theta})$, then multiply the estimator variance by its square.

Check scale and conditions.
Take a square root for the standard error, verify its units, and consider whether smoothness and a first-order approximation are plausible.

The central exam trap is to transform the point estimate correctly but propagate its uncertainty incorrectly. The delta method does not transform a variance by applying $g$ to that variance. Instead, it uses the local slope squared: $[g'(\hat{\theta})]^2\widehat{\operatorname{Var}}(\hat{\theta})$. Another frequent error is to report this variance when the question requests a standard error, or to omit the sample-size factor already present in the variance of $\hat{\theta}$. Keeping the target, derivative, variance, and standard error distinct prevents these errors. This treatment is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Key takeaway: For a smooth scalar function $g$ of an asymptotically normal estimator, approximate the transformed variance by the estimator variance multiplied by the squared local derivative. The result is a first-order approximation whose accuracy depends on the estimator's large-sample behavior and the function's local smoothness.

Check Your Understanding

For each problem, identify the function, use its derivative, and report the requested delta-method quantity. These hypothetical model calculations are intended as independent practice.

Independent exponential losses have mean parameter $\theta$, and the sample mean estimates $\theta$. For $n=25$, $\hat{\theta}=\$1{,}000$; estimate the rate $1/\theta$ and its delta-method standard error. (Answer: rate $0.001$ per dollar; standard error $0.0002$ per dollar.)
For independent lognormal losses, the log standard deviation is known to be $0.5$. With $n=100$, the estimated log mean is $\ln(800)$. Estimate the median and its delta-method standard error. (Answer: median $\$800$; standard error $\$40$.)
A gamma loss model has known shape $k=3$, unknown scale $\theta$, and $n=75$. The observed sample mean is $\$900$. Estimate the model variance and the delta-method standard error of that estimated variance. (Answer: estimated variance $\$270{,}000$ squared; standard error $\$36{,}000$ squared.)
An exponential model's mean is estimated by the sample mean, with $\hat{\theta}=\$2{,}500$ from $n=100$ independent losses. Estimate $g(\theta)=\ln(\theta)$ and its delta-method standard error. (Answer: estimate $\ln(2{,}500)$; standard error $0.1$ on the log scale.)
An estimator has $\hat{\theta}=4$ and estimated standard error $0.3$. For the target $g(\theta)=\theta^2$, calculate the delta-method estimate and standard error. (Answer: estimate $16$; standard error $2.4$.)

What You'll Learn