The Square-Root Transformation for Poisson Rates

Why a Poisson Rate Needs a Different Scale

A Poisson count has variance equal to its mean. Consequently, the variability of a rate estimate depends on the unknown rate itself: rates with larger values have larger standard errors. A transformation can make this changing uncertainty approximately constant, which gives a useful scale for normal approximation and interval construction.

For a Poisson rate, the resulting transformation is a square root. The derivation is simple, but its interpretation matters: the transformed estimate is approximately normal when the total exposure grows and the true rate is positive. It does not make every finite-sample Poisson distribution normal, especially when counts are small.

Poisson Counts and the Rate Estimate

Suppose independent counts \(X_1,\ldots,X_n\) are observed over known exposures \(t_1,\ldots,t_n\), where \(X_i\sim\operatorname{Pois}(t_i\lambda)\) and \(\lambda\geq 0\) is the rate per unit of exposure. Write \(T=\sum_{i=1}^n t_i\) for total exposure and \(Y=\sum_{i=1}^n X_i\) for the total count. Then \(Y\sim\operatorname{Pois}(T\lambda)\), and the rate estimate is \(\hat{\lambda}=Y/T\) when \(T>0\).

The likelihood, up to factors that do not depend on \(\lambda\), is \(L(\lambda)=e^{-T\lambda}\lambda^Y\). For \(Y>0\), differentiating the log-likelihood gives \(Y/\lambda-T=0\), so the maximum occurs at \(\hat{\lambda}=Y/T\). If \(Y=0\), the likelihood is \(e^{-T\lambda}\), which is maximized at \(\lambda=0\) when the parameter space includes zero. If instead the parameter space is restricted to \(\lambda>0\), the likelihood has a supremum as \(\lambda\) decreases to zero but no maximizer. The zero-count case must therefore be treated as a boundary estimate, not as an interior MLE.

Deriving the Square-Root Transformation

For a differentiable function \(g\), the first-order delta method approximates the variance of \(g(\hat{\lambda})\) by \(g'(\lambda)^2\operatorname{Var}(\hat{\lambda})\). To make this approximately independent of \(\lambda\), choose \(g\) so that \(g'(\lambda)^2\lambda\) is constant. Taking the positive root gives a differential equation whose solution is proportional to \(\sqrt{\lambda}\).

The multiplicative constant and additive constant do not affect whether the variance is constant. The conventional choice \(g(\lambda)=2\sqrt{\lambda}\) makes the approximate variance equal to \(1/T\). Equivalently, \(\sqrt{\lambda}\) has approximate variance \(1/(4T)\). We use \(2\sqrt{\lambda}\) below because its standard error has the simple form \(1/\sqrt{T}\).

To justify the result, use \(Y\sim\operatorname{Pois}(T\lambda)\). For fixed \(\lambda>0\), the Poisson central limit theorem gives \((Y-T\lambda)/\sqrt{T\lambda}\xrightarrow{d}N(0,1)\) as \(T\) increases. Since \(\hat{\lambda}=Y/T\), the scalar delta method applied to \(g(u)=2\sqrt{u}\), whose derivative at \(\lambda\) is \(1/\sqrt{\lambda}\), yields

The first-order term equals \((Y-T\lambda)/\sqrt{T\lambda}\), and the remainder is negligible by consistency of \(\hat{\lambda}\) and differentiability of \(g\) at the positive value \(\lambda\). This argument requires a fixed positive rate and increasing total exposure; it does not establish a normal limit at the boundary \(\lambda=0\). \(\square\)

Constructing and Back-Transforming the Interval

Let \(z_{1-\alpha/2}\) be the \(1-\alpha/2\) quantile of the standard normal distribution. On the transformed scale, the approximate two-sided confidence interval for \(2\sqrt{\lambda}\) is centered at \(2\sqrt{\hat{\lambda}}\), with margin \(z_{1-\alpha/2}/\sqrt{T}\). The map \(g(\lambda)=2\sqrt{\lambda}\) is increasing for \(\lambda\geq0\), so its limits can be inverted directly. A negative lower limit on the transformed scale is replaced by zero, since the parameter is nonnegative.

Squaring each endpoint and dividing by four gives the interval on the rate scale:

The interval is approximate, not an exact Poisson interval. Its asymptotic coverage follows from the transformed normal limit: the event that the true transformed parameter lies between the transformed limits is the event that the absolute value of the standardized transformed error is at most \(z_{1-\alpha/2}\). The limiting probability of this event is \(1-\alpha\). Since the transformation is increasing and the true rate is nonnegative, clipping a negative lower endpoint to zero does not remove the true parameter from the interval.

Interpretation and Common Pitfalls

The square-root transformation is useful because it addresses the rate dependence of the first-order variance. It gives a natural approximate interval whose lower endpoint cannot be negative and whose uncertainty is measured on a scale with approximately constant variance. It is not a correction that makes every Poisson problem normal, nor is it an exact confidence procedure.

• Using the untransformed standard error after transforming. The original-scale standard error \(\sqrt{\hat{\lambda}/T}\) is not the standard error of \(2\sqrt{\hat{\lambda}}\). On the transformed scale, the first-order standard error is \(1/\sqrt{T}\).
• Forgetting the inverse transformation. The transformed endpoints are endpoints for \(2\sqrt{\lambda}\), not for \(\lambda\). Invert by squaring and dividing by four.
• Claiming an interior MLE at zero counts. If \(\lambda>0\) is imposed, \(Y=0\) yields no attained maximum. Include \(\lambda=0\) to call zero the MLE, or describe it as the boundary supremum estimate under the restricted parameter space.
• Assuming asymptotic validity ensures good small-sample coverage. The normal limit is for fixed \(\lambda>0\) with increasing exposure. A small observed count can be a warning that the approximation is rough.

A useful check is to verify units and scale: \(\hat{\lambda}\) is a rate, \(T\) is exposure, and the transformed-scale margin depends on \(T\) as \(1/\sqrt{T}\). After inversion, both interval endpoints are nonnegative and expressed in the same rate units as \(\hat{\lambda}\).