Delta-Method Uncertainty for Price Elasticities

Why Elasticity Uncertainty Needs Its Own Calculation

A fitted demand model may estimate its coefficients precisely while leaving substantial uncertainty in a particular price elasticity. The elasticity is often a nonlinear function of several coefficients, and its uncertainty depends not just on their individual standard errors but also on how their estimates move together. A coefficient standard error by itself is therefore not an elasticity standard error.

The delta method provides a first-order approximation to the sampling variance of a smooth function of estimated coefficients. It applies to probability elasticities from logistic models and to elasticities from nonlinear transformations of demand curves, provided the relevant function is differentiable near the fitted coefficients and the coefficient estimator has a usable covariance estimate.

Definition: Let $\boldsymbol{\beta}$ be a fixed coefficient vector, let $\hat{\boldsymbol{\beta}}$ be its estimator, and let $\widehat{\mathbf V}$ estimate $\operatorname{Cov}(\hat{\boldsymbol{\beta}})$. For a differentiable scalar function $g$, the delta-method variance estimate for $g(\hat{\boldsymbol{\beta}})$ is $$ \widehat{\operatorname{Var}}\{g(\hat{\boldsymbol{\beta}})\} = \nabla g(\hat{\boldsymbol{\beta}})^{\mathsf T} \widehat{\mathbf V} \nabla g(\hat{\boldsymbol{\beta}}), $$ where $\nabla g$ is the column vector of partial derivatives with respect to the coefficients. The approximate standard error is the square root of this variance estimate.

The covariance matrix must use the same coefficient order as the gradient. It includes both coefficient variances and, when present, covariances between coefficients. The price and any other covariates at which elasticity is evaluated are held fixed in this calculation; the target is uncertainty from estimating the model coefficients.

Why the Delta-Method Formula Works

Suppose the coefficient estimator is approximately normal in large samples, with mean near the fixed true vector $\boldsymbol{\beta}$ and covariance $\mathbf V$. A first-order Taylor expansion around that true vector gives

g(\hat{\boldsymbol{\beta}}) \approx g(\boldsymbol{\beta}) + \nabla g(\boldsymbol{\beta})^{\mathsf T} (\hat{\boldsymbol{\beta}}-\boldsymbol{\beta}).

Here $g(\boldsymbol{\beta})$ and $\nabla g(\boldsymbol{\beta})$ are fixed. The random quantity is the estimator $\hat{\boldsymbol{\beta}}$. Therefore, the approximate variance comes from the linear term:

\operatorname{Var}\{g(\hat{\boldsymbol{\beta}})\} \approx \nabla g(\boldsymbol{\beta})^{\mathsf T} \operatorname{Cov}(\hat{\boldsymbol{\beta}}) \nabla g(\boldsymbol{\beta}).

The true coefficients and their covariance are unknown, so estimation substitutes $\hat{\boldsymbol{\beta}}$ and $\widehat{\mathbf V}$. Under the additional approximation that the transformed estimator is normally distributed, a two-sided 95% Wald interval is the estimate plus or minus $1.96$ standard errors. This is an approximation, not a guarantee of exact coverage; strong curvature or a highly skewed elasticity estimator can make it less reliable.

Logistic Probability Elasticity

Consider a binary purchase outcome with a logistic model

p(P)=\frac{1}{1+\exp(-\eta)}, \qquad \eta=a+b x, \qquad x=\ln(P/P_0),

where $p(P)$ is purchase probability and $P_0$ is a fixed reference price. The probability elasticity is the percentage change in purchase probability associated with a percentage change in price: $\varepsilon=(1/p)(dp/d\ln P)$. Since $dx/d\ln P=1$ and $dp/d\eta=p(1-p)$, it follows that

\varepsilon(a,b;x)=b(1-p).

This is an elasticity of probability, not the coefficient $b$ itself. Because $p$ also depends on $a$ and $b$, the elasticity is a nonlinear function of both coefficients. Differentiating gives

\nabla\varepsilon = \begin{pmatrix} -bp(1-p)\\ (1-p)-b x p(1-p) \end{pmatrix}.

The first component follows from $\partial p/\partial a=p(1-p)$. For the second, apply the product rule to $b(1-p)$, using $\partial p/\partial b=xp(1-p)$.

Worked Example: Uncertainty at the Reference Price

A hypothetical insurer models purchase probability as logistic in $x=\ln(P/P_0)$. At the fitted coefficients, $a=0$ and $b=-1$. At $P=P_0$, calculate the probability elasticity, its delta-method standard error, and an approximate 95% interval. Assume the coefficient estimator is approximately normal and its estimated covariance matrix, in the order $(a,b)$, is diagonal with variances $0.04$ and $0.01$. No interest or mortality assumptions enter this demand calculation.

Method. Evaluate the elasticity and its gradient at the fitted coefficients and target price, then apply the covariance-matrix formula. The target has $x=\ln(1)=0$, so $\eta=0+(-1)(0)=0$, and $p=1/(1+\exp(0))=0.5$.

The fitted elasticity is $\hat{\varepsilon}=(-1)(1-0.5)=-0.5$. The gradient is $\nabla\varepsilon=(-b p(1-p),(1-p)-b x p(1-p))^{\mathsf T}=(0.25,0.5)^{\mathsf T}$.

The covariance matrix is $\widehat{\mathbf V}=\begin{pmatrix}0.04&0\\0&0.01\end{pmatrix}$. Thus

\widehat{\operatorname{Var}}(\hat{\varepsilon}) = 0.25^2(0.04)+0.5^2(0.01) = 0.0025+0.0025 = 0.0050, \qquad \operatorname{SE}(\hat{\varepsilon})=\sqrt{0.0050}=0.0707.

The approximate 95% interval is $-0.5\pm1.96(0.0707107)=-0.5\pm0.1386$, or $[-0.6386,-0.3614]$, to four decimals. This method avoids treating the probability elasticity as though it were simply $b$; doing so would miss the dependence of $p$ on both fitted coefficients.

Using Covariances Between Coefficients

For a two-coefficient vector with gradient $(g_1,g_2)^{\mathsf T}$ and covariance matrix $\begin{pmatrix}V_{11}&V_{12}\\V_{12}&V_{22}\end{pmatrix}$, the variance calculation expands to

g_1^2V_{11}+2g_1g_2V_{12}+g_2^2V_{22}.

The middle term can increase or decrease the result, depending on the signs of the gradient components and covariance. Omitting it is valid only when the estimated covariance is actually zero.

Worked Example: Logistic Elasticity With Correlated Estimates

For another hypothetical binary-demand model, let $a=0$, $b=-0.5$, and evaluate at $x=1$. Assume approximate normality of the coefficient estimator and estimated covariance matrix $\widehat{\mathbf V}=\begin{pmatrix}0.01&-0.001\\-0.001&0.0004\end{pmatrix}$, in the order $(a,b)$. Calculate the elasticity, its standard error, and its approximate 95% interval. The covariance matrix is admissible: its determinant is $0.01(0.0004)-0.001^2=0.000003>0$.

Method. First evaluate the probability: $\eta=0+(-0.5)(1)=-0.5$, so $p=1/(1+\exp(0.5))=0.3775$ to four decimals. Using the unrounded value $p=0.3775407$, $p(1-p)=0.2350037$, and $\hat{\varepsilon}=(-0.5)(1-0.3775407)=-0.3112297$.

The gradient components, evaluated without intermediate rounding, are $g_1=-b p(1-p)=0.1175019$ and $g_2=(1-p)-b x p(1-p)=0.6224593+0.1175019=0.7399612$. Including all covariance terms,

\begin{aligned} \widehat{\operatorname{Var}}(\hat{\varepsilon}) &=g_1^2(0.01)+2g_1g_2(-0.001)+g_2^2(0.0004)\\ &\approx 0.00013807-0.00017389+0.00021902\\ &\approx 0.00018319, \qquad \operatorname{SE}(\hat{\varepsilon})\approx 0.01353. \end{aligned}

The interval is $-0.3112297\pm1.96(0.0135348)$, approximately $[-0.3378,-0.2847]$, to four decimals. A common incorrect calculation drops the covariance term, producing variance approximately $0.00013807+0.00021902=0.00035709$ and standard error $0.0189$. The correction is to use the full matrix: the negative covariance contributes $2g_1g_2(-0.001)$, reducing the variance in this case.

Elasticity From a Polynomial Demand Curve

The same procedure applies when demand is modeled directly as a function of price. For a positive fitted demand quantity $q(P)=a+bP+cP^2$, the price elasticity is $\varepsilon=(P/q)(dq/dP)$. Since $dq/dP=b+2cP$, define $N=P(b+2cP)$, so that $\varepsilon=N/q$. Differentiation with respect to the coefficient vector $(a,b,c)$ gives

\nabla\varepsilon = \begin{pmatrix} -\dfrac{N}{q^2}\\[4pt] \dfrac{P}{q}-\dfrac{NP}{q^2}\\[4pt] \dfrac{2P^2}{q}-\dfrac{NP^2}{q^2} \end{pmatrix}.

These derivatives follow from the quotient rule: the numerator has coefficient derivatives $0,P,2P^2$, while the denominator has derivatives $1,P,P^2$. The calculation requires $q(P)\ne0$; for an interpretable demand elasticity, the modeled demand should be positive at the target price.

Worked Example: Polynomial-Demand Elasticity and Interval

A hypothetical product has fitted demand $q(P)=a+bP+cP^2$, with estimates $a=20$, $b=-4$, and $c=0.5$. At price $P=2$, calculate the elasticity, delta-method standard error, and approximate 95% interval. Assume approximate normality and a diagonal estimated coefficient covariance matrix with variances $0.25$, $0.01$, and $0.0004$, respectively. The price and demand units are held fixed for this calculation; no interest or mortality assumption is relevant.

Method. At $P=2$, fitted demand is $q=20+(-4)(2)+0.5(2^2)=20-8+2=14$. The numerator is $N=P(b+2cP)=2[-4+2(0.5)(2)]=2(-2)=-4$. Therefore, $\hat{\varepsilon}=N/q=-4/14=-2/7=-0.2857143$.

The gradient components are $-N/q^2=4/196=1/49$, $P/q-NP/q^2=2/14-(-4)(2)/196=1/7+2/49=9/49$, and $2P^2/q-NP^2/q^2=8/14-(-4)(4)/196=4/7+4/49=32/49$. Thus the gradient is $(1/49,9/49,32/49)^{\mathsf T}$. With a diagonal covariance matrix, the variance is

\begin{aligned} \widehat{\operatorname{Var}}(\hat{\varepsilon}) &=(1/49)^2(0.25)+(9/49)^2(0.01)+(32/49)^2(0.0004)\\ &=\frac{0.25+0.81+0.4096}{2401} =\frac{1.4696}{2401} \approx 0.00061208,\\ \operatorname{SE}(\hat{\varepsilon})&\approx\sqrt{0.00061208}=0.02474. \end{aligned}

The approximate 95% interval is $-0.2857143\pm1.96(0.0247402)=[-0.3342,-0.2372]$, to four decimals. This calculation avoids the mistake of using only the uncertainty in the slope $b$: the elasticity also changes with the intercept and quadratic coefficient through both $N$ and $q$.

Practical Interpretation and Exam Traps

A delta-method interval quantifies local sampling uncertainty in a fitted elasticity at a specified price and covariate profile. It does not establish that the demand curve is well calibrated, that an extrapolated price is supported by the data, or that a local approximation remains accurate far from the fitted point. Those are separate model-assessment questions.

Use the estimator's covariance. In the frequentist derivation, the coefficients are fixed and the estimator varies across samples. Expanding around the true coefficient vector gives the variance formula; it is not an argument that treats the true coefficients as random conditional on a fitted estimate.
Differentiate the full elasticity function. For logistic purchase probability, $\varepsilon=b(1-p)$, not simply $b$. The factor $1-p$ changes with the fitted coefficients.
Retain covariance terms. A diagonal-only calculation is wrong when off-diagonal entries are nonzero.
Match ordering and units. Reorder the gradient if the covariance matrix uses a different coefficient order. A price transformation such as $x=\ln(P/P_0)$ must also be differentiated consistently when deriving the elasticity.
Interpret an approximate interval carefully. A symmetric Wald interval can be a poor approximation if the elasticity function is highly curved over the plausible coefficient range.

For an exam calculation, a reliable sequence is: define the elasticity as $g(\boldsymbol{\beta})$, evaluate it at the fitted coefficients, derive its gradient, substitute the estimated covariance matrix, take the square root, and only then construct the requested interval. This sequence is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

For each problem, use the stated coefficient order and retain full precision until the final answer. Intervals are approximate 95% Wald intervals.

A logistic purchase model has $\eta=a+bx$, with estimates $a=0$, $b=-1$, evaluated at $x=0$. The estimated covariance matrix for $(a,b)$ is diagonal with variances $0.04$ and $0.01$. Find the delta-method standard error of the probability elasticity. (Answer: $0.0707$)
A logistic model has $\eta=a+bx$, with estimates $a=0$, $b=-0.5$. At $x=1$, calculate the probability elasticity. (Answer: approximately $-0.3112$)
A hypothetical linear demand curve is $q(P)=a+bP$, with $a=20$, $b=-2$, evaluated at $P=4$. The estimated covariance matrix for $(a,b)$ is diagonal with variances $0.09$ and $0.0009$. Find the elasticity's delta-method standard error. (Answer: $0.0236$)
For $q(P)=a+bP+cP^2$, use $a=20$, $b=-4$, $c=0.5$, and $P=2$. The estimated coefficient covariance matrix is diagonal with variances $0.25$, $0.01$, and $0.0004$. Find the approximate 95% interval for elasticity. (Answer: approximately $[-0.3342,-0.2372]$)
For a logistic model, use $a=0$, $b=-0.5$, and $x=1$. The estimated covariance matrix for $(a,b)$ is $\begin{pmatrix}0.01&-0.001\\-0.001&0.0004\end{pmatrix}$. Find the delta-method standard error and approximate 95% interval for probability elasticity. (Answer: standard error $0.0135$; interval approximately $[-0.3378,-0.2847]$)

What You'll Learn