Choosing a Transformation From an Asymptotic Variance Function

Choosing a Scale From the Variance Function

A variance-stabilizing transformation is not chosen by looking for a familiar formula first. The starting point is the estimator’s asymptotic variance: which smooth change of scale makes its first-order variability approximately constant across the parameter space? Once that transformation is derived, its domain, monotonicity, and interpretation still need to be checked. A formula that solves the variance equation may be mathematically valid while being awkward near a boundary or difficult to explain in the application.

Suppose an estimator \(\hat{\theta}\) targets a scalar parameter \(\theta\) in an interval \(\Theta\), and its first-order asymptotic variance is \(V(\theta)/n\). For the derivation below, assume that \(\theta\) is an interior point of \(\Theta\), \(V\) is continuous and strictly positive on the interval under consideration, and the estimator has the asymptotic distribution

\sqrt{n}(\hat{\theta}-\theta)\xrightarrow{d}N(0,V(\theta)).

This distributional assumption supplies both the \(n^{-1/2}\) scale and the local normal limit needed for the delta method. A variance function by itself does not establish this limit.

Definition: A differentiable transformation \(g\) is first-order variance-stabilizing for \(V\) on an interval \(\Theta\) if \(g'(\theta)^2V(\theta)\) is constant and positive throughout that interval. Under the stated asymptotic distribution, this makes the first-order asymptotic variance of \(g(\hat{\theta})\) constant as a function of \(\theta\).

Deriving and Checking the Transformation

For a differentiable \(g\) with \(g'(\theta)\ne 0\), the delta method gives

\sqrt{n}\{g(\hat{\theta})-g(\theta)\} \xrightarrow{d} N\left(0,g'(\theta)^2V(\theta)\right).

Thus, to make the limiting variance equal to one, solve \(g'(\theta)^2V(\theta)=1\). On a connected interval where \(V\) is positive, a continuous derivative cannot change sign without passing through zero. The increasing choice is therefore \(g'(\theta)=1/\sqrt{V(\theta)}\); the decreasing choice is its negative. Integrating gives the usual construction.

Construction: If \(V(\theta)>0\) on an interval, a unit-variance first-order scale is

g(\theta)=\int_{\theta_0}^{\theta}\frac{du}{\sqrt{V(u)}},

where \(\theta_0\) is any fixed point in the interval. Adding a constant to \(g\), or multiplying it by \(-1\), leaves its derivative squared unchanged. More generally, a nonzero constant multiplier changes the stabilized variance to that multiplier squared; it does not destroy constancy.

Theorem: Let \(\Theta\) be an interval and let \(V\) be continuous and strictly positive there. If \(g\) is differentiable and \(g'(\theta)^2V(\theta)=C\) for a fixed \(C>0\), then on \(\Theta\), \(g\) has one of the forms

g(\theta)=a\ \mathord{\pm}\ \sqrt{C}\int_{\theta_0}^{\theta}\frac{du}{\sqrt{V(u)}},

for a constant \(a\). Conversely, each such function has \(g'(\theta)^2V(\theta)=C\).

Proof. The equation implies \(g'(\theta)^2=C/V(\theta)\), so \(g'(\theta)\) is never zero. Since \(g'\) has the intermediate value property, its sign is constant on the interval: if it took both signs, it would equal zero somewhere between those points. Therefore \(g'(\theta)=\sqrt{C}/\sqrt{V(\theta)}\) throughout the interval, or \(g'(\theta)=-\sqrt{C}/\sqrt{V(\theta)}\) throughout. Integrating from \(\theta_0\) to \(\theta\) gives the displayed forms, with \(a=g(\theta_0)\). Differentiating either form and squaring verifies the converse. \(\square\)

Identify the variance function.
Write the asymptotic variance as \(V(\theta)/n\), separating out the sample-size factor.

Check the domain.
Find the parameter interval where \(V\) is finite, continuous, and strictly positive.

Integrate the reciprocal square root.
Use \(g'(\theta)=1/\sqrt{V(\theta)}\) for an increasing scale, and choose any convenient integration constant.

Verify the proposed scale.
Differentiate it, check that its derivative has one sign, and calculate \(g'(\theta)^2V(\theta)\).

Assess the result.
Check the range, endpoint behavior, and whether differences on the transformed scale can be explained usefully in the application.

The theorem is a local, first-order calculation on an interval where \(V\) is positive. If \(V\) vanishes or diverges at a boundary, the integral may still have a finite or infinite limit there, but differentiability and the original asymptotic approximation need to be considered at interior parameter values. Do not assume that a formula valid in the interior automatically gives a well-behaved inference procedure at a boundary.

Worked Example: Choosing a Scale for a Binomial Proportion

Let \(\hat{p}=X/n\), where \(X\sim\operatorname{Binomial}(n,p)\) and \(0<p<1\). Its variance is \(p(1-p)/n\), so \(V(p)=p(1-p)\). Find a unit-variance first-order transformation and check it.

For an increasing transformation, integrate the reciprocal square root:

g(p)=\int_0^p\frac{du}{\sqrt{u(1-u)}}.

Use \(u=\sin^2(t)\). On \(0<t<\pi/2\), \(du=2\sin(t)\cos(t)\,dt\) and \(\sqrt{u(1-u)}=\sin(t)\cos(t)\). The integral is therefore \(2t\), giving

g(p)=2\arcsin(\sqrt{p}).

Direct differentiation verifies the result:

g'(p)=\frac{1}{\sqrt{p}\sqrt{1-p}}, \qquad g'(p)^2V(p) =\frac{1}{p(1-p)}p(1-p)=1.

The derivative is positive for \(0<p<1\), so the transformation is strictly increasing there. It extends continuously to the endpoints with \(g(0)=0\) and \(g(1)=\pi\), although the derivative is unbounded near them. Its range is \((0,\pi)\) for interior proportions. The transformed scale is bounded and orders proportions in the same way as the original scale; its distances are not percentage-point differences.

Worked Example: Choosing a Scale for a Poisson Rate

Suppose \(X_1,\ldots,X_n\) are independent Poisson observations with mean \(\theta>0\), and let \(\hat{\theta}=\bar{X}\). Since \(\operatorname{Var}(\bar{X})=\theta/n\), the variance function is \(V(\theta)=\theta\). Find and check the increasing transformation.

Integrating gives

g(\theta)=\int_0^\theta\frac{du}{\sqrt{u}}=2\sqrt{\theta}.

For \(\theta>0\), \(g'(\theta)=1/\sqrt{\theta}>0\), and hence

g'(\theta)^2V(\theta) =\frac{1}{\theta}\theta=1.

The transformed parameter is \(2\sqrt{\theta}\), which preserves the ordering of positive rates but expresses differences on a square-root scale. For example, a one-unit increase on this scale corresponds to a change in \(\theta\) that depends on its starting value, since \(\theta=g(\theta)^2/4\). The transformation extends continuously to zero, but its derivative is unbounded there. The interior derivation alone does not establish a normal approximation at \(\theta=0\).

Worked Example: A Smooth Variance Function on the Real Line

Consider an estimator with asymptotic distribution \(\sqrt{n}(\hat{\theta}-\theta)\xrightarrow{d}N(0,1+\theta^2)\), where \(\theta\) ranges over all real numbers. Here \(V(\theta)=1+\theta^2\), which is continuous and strictly positive everywhere.

The increasing transformation is

g(\theta)=\int_0^\theta\frac{du}{\sqrt{1+u^2}} =\operatorname{arsinh}(\theta) =\log\left(\theta+\sqrt{1+\theta^2}\right).

Its derivative is \(g'(\theta)=1/\sqrt{1+\theta^2}>0\), and direct substitution gives

g'(\theta)^2V(\theta) =\frac{1}{1+\theta^2}(1+\theta^2)=1.

The function is strictly increasing and maps the real line onto the real line. Near zero, \(\operatorname{arsinh}(\theta)\) is approximately \(\theta\); for large positive \(\theta\), it grows approximately like \(\log(2\theta)\). Thus it preserves order while compressing large positive values. These features may be helpful for representing a broad range of estimates, but whether that scale is meaningful depends on the scientific units and the audience.

Mathematical Validity and Practical Interpretation

The transformation solves a specific mathematical problem: it makes the first-order asymptotic variance constant. It does not guarantee that the transformed estimator is exactly normal in finite samples, that its distribution is symmetric, or that a normal approximation is accurate near a parameter boundary. In the examples, the delta-method conclusion follows from the stated asymptotic normality at an interior parameter value; the variance calculation alone would not justify the distributional conclusion.

A strictly monotone transformation preserves the order of parameter values, which is often useful for interpretation and for translating intervals back to the original scale. But transformed-scale distances have their own units and meaning. In the Poisson example, for instance, equal changes in \(2\sqrt{\theta}\) do not correspond to equal changes in \(\theta\). Ask whether the transformed scale is communicable and whether its endpoint behavior is acceptable for the intended analysis. Choosing the increasing rather than decreasing version usually makes the scale easier to read; adding a constant can also place its origin conveniently without changing its variance-stabilizing property.

Common mistake: Choosing \(g'(\theta)=1/V(\theta)\) because the aim is to “divide out the variance.” The delta method multiplies the variance by \(g'(\theta)^2\), not by \(g'(\theta)\). The required equation is \(g'(\theta)^2V(\theta)=C\), so the derivative must be proportional to \(1/\sqrt{V(\theta)}\). For example, with \(V(\theta)=\theta\), choosing \(g'(\theta)=1/\theta\) gives \(g'(\theta)^2V(\theta)=1/\theta\), which is not constant.

The practical choice is therefore not simply “integrate and stop.” First verify that the variance function is positive and regular on the parameter region of interest; then derive and differentiate the candidate; finally inspect its monotonicity, range, and behavior at relevant boundaries. If multiple transformations differ only by a location or scale constant, those choices can be guided by a convenient origin and a meaningful unit. None of these decisions replaces checking the estimator’s actual sampling behavior.

Check Your Understanding

Use the variance-function construction and distinguish its mathematical conclusion from claims about finite-sample behavior.

If \(V(\theta)=4\theta^2\) for \(\theta>0\), find an increasing unit-variance transformation and verify it by differentiation.
Why does a positive continuous variance function on an interval imply that a transformation with \(g'(\theta)^2V(\theta)=C>0\) has a derivative of constant sign?
For the binomial transformation, what is its range for \(0<p<1\), and why does that not mean its derivative remains finite at the endpoints?
What additional distributional condition, beyond knowing \(V(\theta)/n\), supports the stated delta-method normal limit?
Give one reason a mathematically valid variance-stabilizing scale might be inconvenient to interpret in a particular application.

Choosing a Transformation From an Asymptotic Variance Function

What You'll Learn