Delta-Method Intervals for Ratios

Why Ratios Need Special Care

A ratio can summarize a relative effect, a rate, or a change per unit of another quantity. Its calculation is simple, but its uncertainty can be sensitive to the denominator: when the denominator is close to zero, small changes in either estimate may produce large changes in the ratio. A standard large-sample interval can be useful when the denominator is well separated from zero, but the conditions behind that approximation matter.

This tutorial develops a delta-method interval for the ratio of two jointly estimated quantities. The covariance between those estimates is part of the calculation; treating them as independent when they are not can misstate the standard error. The interval is a local, approximately normal procedure, so we will also examine whether the denominator estimate is sufficiently far from zero to make that procedure credible.

Definition: Let \(\hat a\) and \(\hat b\) estimate parameters \(a\) and \(b\), respectively, with \(b\ne 0\). The ratio estimand is \(\psi=g(a,b)=a/b\), and its plug-in estimator is \(\hat\psi=\hat a/\hat b\), provided \(\hat b\ne 0\). A delta-method interval approximates the sampling distribution of \(\hat\psi\) by linearizing \(g\) around \((a,b)\) and estimating the resulting variance.

The Gradient and the Ratio Variance

The gradient records how the ratio changes when either component changes. Differentiating \(g(a,b)=a/b\) gives

\nabla g(a,b)= \begin{pmatrix} 1/b\\ -a/b^2 \end{pmatrix}.

Suppose the estimates are jointly asymptotically normal and their estimated covariance matrix is available. Write

\hat{\boldsymbol\eta}= \begin{pmatrix}\hat a\\\hat b\end{pmatrix}, \qquad \hat V= \begin{pmatrix} \widehat{\operatorname{Var}}(\hat a)& \widehat{\operatorname{Cov}}(\hat a,\hat b)\\ \widehat{\operatorname{Cov}}(\hat a,\hat b)& \widehat{\operatorname{Var}}(\hat b) \end{pmatrix}.

The estimated delta-method variance evaluates the gradient at the estimates and propagates their joint uncertainty through the ratio:

\hat v_\psi = \nabla g(\hat a,\hat b)^\mathsf{T} \hat V \nabla g(\hat a,\hat b) = \frac{\widehat{\operatorname{Var}}(\hat a)}{\hat b^2} + \frac{\hat a^2}{\hat b^4}\widehat{\operatorname{Var}}(\hat b) - \frac{2\hat a}{\hat b^3} \widehat{\operatorname{Cov}}(\hat a,\hat b).

The last term is not optional: it represents the dependence between the two estimates. It can increase or decrease the variance, depending on the sign of their covariance.

Theorem: Suppose \(b\ne0\), \(\sqrt n\{(\hat a,\hat b)^\mathsf{T}-(a,b)^\mathsf{T}\}\xrightarrow{d}N(0,\Sigma)\), and \(n\hat V\xrightarrow{p}\Sigma\), where \(\Sigma\) is a finite covariance matrix. If \(v=\nabla g(a,b)^\mathsf{T}\Sigma\nabla g(a,b)>0\), then \(\sqrt n(\hat\psi-\psi)\xrightarrow{d}N(0,v)\). Moreover, if \(\hat v_\psi\) is defined by the plug-in formula above, then \(n\hat v_\psi\xrightarrow{p}v\), and the studentized statistic converges in distribution to \(N(0,1)\). Thus, for \(0<\alpha<1\), an approximate two-sided \(100(1-\alpha)\%\) interval is \(\hat\psi\pm z_{1-\alpha/2}\sqrt{\hat v_\psi}\), provided the estimated variance is positive.

The condition \(v>0\) is essential: a positive estimated variance at a particular sample size does not by itself establish the stated asymptotic result. In applications, the interval also requires \(\hat b\ne0\) so that the ratio and plug-in gradient are defined.

Why the Interval Has Its Stated Coverage

The justification uses a first-order Taylor expansion, followed by consistency of the estimated covariance and Slutsky’s theorem. The expansion must be applied where the ratio is differentiable; this is why the population denominator must be nonzero.

Linearize the ratio.
Since \(b\ne0\), \(g(a,b)=a/b\) is differentiable in a neighborhood of \((a,b)\). Consistency of \(\hat{\boldsymbol\eta}\) implies that it lies in this neighborhood with probability tending to one. The first-order Taylor expansion is \(\hat\psi-\psi=\nabla g(a,b)^\mathsf{T}(\hat{\boldsymbol\eta}-\boldsymbol\eta)+r_n\), where \(r_n=o_p(\|\hat{\boldsymbol\eta}-\boldsymbol\eta\|)\).

Obtain the limiting distribution.
The assumed joint central limit theorem gives \(\sqrt n(\hat{\boldsymbol\eta}-\boldsymbol\eta)=O_p(1)\), so \(\sqrt n r_n=o_p(1)\). The linear term converges to a centered normal variable with variance \(\nabla g(a,b)^\mathsf{T}\Sigma\nabla g(a,b)=v\). Therefore \(\sqrt n(\hat\psi-\psi)\xrightarrow{d}N(0,v)\).

Estimate and studentize the variance.
Consistency of \(\hat{\boldsymbol\eta}\), continuity of the gradient where \(b\ne0\), and \(n\hat V\xrightarrow{p}\Sigma\) give \(n\hat v_\psi\xrightarrow{p}v\). Since \(v>0\), \(\sqrt{n\hat v_\psi}\xrightarrow{p}\sqrt v\). Slutsky’s theorem then gives \((\hat\psi-\psi)/\sqrt{\hat v_\psi}\xrightarrow{d}N(0,1)\).

Translate the limit into coverage.
For \(z_{1-\alpha/2}\), the \(1-\alpha/2\) quantile of the standard normal distribution, the limiting probability that the studentized statistic lies between \(-z_{1-\alpha/2}\) and \(z_{1-\alpha/2}\) is \(1-\alpha\). This is the coverage claim for the stated interval. \(\square\)

Worked Examples

Worked Example: Ratio of Two Independent Proportions

In two independent groups, 72 of 120 participants experience an outcome in group 1, and 48 of 100 do so in group 2. Estimate the ratio of the two proportions and form a 95% delta-method interval.

The estimates are \(\hat a=72/120=0.60\) and \(\hat b=48/100=0.48\), so \(\hat\psi=0.60/0.48=1.25\). For independent Bernoulli samples, the estimated variances are \(0.60(0.40)/120=0.002000\) and \(0.48(0.52)/100=0.002496\), and the estimated covariance is zero. The gradient at the estimates is \((1/0.48,\,-0.60/0.48^2)^\mathsf{T}=(2.08333,\,-2.60417)^\mathsf{T}\).

Thus the estimated variance is \(0.002000/0.48^2+(0.60^2)(0.002496)/0.48^4=0.0086806+0.0169259=0.0256065\), giving a standard error of \(\sqrt{0.0256065}=0.1600\). Using \(z_{0.975}=1.96\), the interval is \(1.25\pm1.96(0.1600)\), or approximately \((0.936,\;1.564)\).

The denominator estimate has standard error \(\sqrt{0.002496}=0.04996\). Its ordinary Wald interval is \(0.48\pm1.96(0.04996)=(0.382,\;0.578)\), which excludes zero. This check supports using a local ratio approximation here; it is a diagnostic, not a guarantee of good finite-sample coverage.

Worked Example: Including Covariance in a Ratio Interval

Suppose two estimates are \(\hat a=2.1\) and \(\hat b=3.9\), and their estimated covariance matrix is \(\hat V=\begin{pmatrix}0.04&0.012\\0.012&0.09\end{pmatrix}\). Construct the 95% delta-method interval.

The ratio estimate is \(2.1/3.9=0.53846\). The plug-in gradient is \((1/3.9,\,-2.1/3.9^2)^\mathsf{T}=(0.256410,\,-0.138067)^\mathsf{T}\). Using the full covariance matrix gives

\hat v_\psi =(0.256410)^2(0.04) +(-0.138067)^2(0.09) +2(0.256410)(-0.138067)(0.012) \approx 0.003496.

The standard error is \(\sqrt{0.003496}\approx0.05913\). The interval is \(0.53846\pm1.96(0.05913)\), approximately \((0.423,\;0.654)\). If the covariance term were incorrectly omitted, the estimated variance would instead be about \(0.004345\), producing a larger standard error. Here the positive covariance reduces the ratio variance because the two gradient components have opposite signs.

The denominator estimate has standard error \(\sqrt{0.09}=0.3\), so its Wald interval is \(3.9\pm1.96(0.3)=(3.312,\;4.488)\). It is well separated from zero on this scale.

Worked Example: A Denominator Estimate Near Zero

Suppose \(\hat a=1.2\), \(\hat b=0.08\), and the estimated covariance matrix is \(\hat V=\begin{pmatrix}0.01&0\\0&0.01\end{pmatrix}\). Calculate the usual delta-method interval and assess the denominator.

The estimated ratio is \(1.2/0.08=15\), and the gradient is \((12.5,\,-187.5)^\mathsf{T}\). The estimated variance is \(12.5^2(0.01)+187.5^2(0.01)=1.5625+351.5625=353.125\), so the standard error is approximately \(18.79\). The nominal 95% interval is \(15\pm1.96(18.79)\), approximately \((-21.8,\;51.8)\).

But the denominator standard error is \(0.1\), and its Wald interval is \(0.08\pm1.96(0.1)=(-0.116,\;0.276)\), which includes zero. The ratio estimate is therefore highly sensitive to plausible changes in the denominator. The large-sample theorem describes behavior under its stated fixed-parameter assumptions; this calculation does not, by itself, establish that those assumptions give a reliable finite-sample approximation here. A ratio Wald interval should not be treated as reassuring merely because it can be computed.

Assessing Denominator Separation

For a fixed target with \(b\ne0\) and a consistent estimator \(\hat b\), the denominator eventually stays away from zero with probability tending to one. In finite samples, however, it is useful to compare \(\hat b\) with its standard error. One simple diagnostic is the interval \(\hat b\pm z_{1-\alpha/2}\sqrt{\widehat{\operatorname{Var}}(\hat b)}\). If that interval crosses zero, a local approximation centered at the estimated denominator may be unstable. If it does not cross zero, the denominator is more clearly separated on the scale of its estimated uncertainty, though this alone does not guarantee accurate coverage for the ratio.

A very small standard error for the ratio is not the only criterion that matters: the approximation relies on a smooth, locally linear transformation near the true parameter. When denominator uncertainty is substantial relative to its distance from zero, confidence procedures designed for uncertain denominators may be more appropriate. Those procedures can yield confidence sets that are not ordinary bounded intervals.

Common Mistakes

Ignoring covariance. Setting \(\widehat{\operatorname{Cov}}(\hat a,\hat b)=0\) without justification drops a term that may materially change the standard error. Use the joint covariance estimate unless independence supports a zero covariance.
Checking only that the estimated variance is positive. Positivity at the observed sample size is not enough for the studentized asymptotic result. The limiting delta-method variance \(v=\nabla g(a,b)^\mathsf{T}\Sigma\nabla g(a,b)\) must also be positive.
Assuming a nonzero observed denominator guarantees stability. The ratio can be defined while the denominator remains poorly separated from zero relative to its uncertainty. Inspect its scale and uncertainty, not only whether \(\hat b\ne0\).
Using the interval as a universal ratio procedure. The delta-method interval is a local normal approximation. Its usefulness depends on the denominator and the quality of the underlying large-sample approximations.

Key takeaway: For a ratio, evaluate the gradient \((1/b,\,-a/b^2)^\mathsf{T}\) against the full covariance matrix of the numerator and denominator estimates. Use the resulting Wald interval only under a positive limiting delta-method variance, and assess whether the denominator is separated from zero relative to its uncertainty.

Check Your Understanding

Answer each question by identifying the relevant gradient, covariance, or denominator condition.

For \(\psi=a/b\), derive the two components of \(\nabla g(a,b)\).
In the ratio variance formula, what determines whether the covariance contribution increases or decreases the variance?
Why does a positive estimated variance at the observed sample size not alone establish the studentized asymptotic result?
For \(\hat a=4\), \(\hat b=2\), and estimated variances \(0.16\) and \(0.09\) with estimated covariance \(0.03\), write the plug-in delta-method variance expression.
What does it suggest if the denominator’s Wald interval includes zero, and why is that a diagnostic rather than a definitive coverage test?