Saddlepoint Calibration for Imbalanced Binary Traits

Why Calibrate a Score-Test Tail?

In a binary genetic association analysis, a score test evaluates whether a genetic effect is compatible with the null model, often a logistic regression model containing covariates. Its usual large-sample reference distribution is a standard normal distribution for a standardized score, or a chi-square distribution for the squared score. With highly imbalanced case-control counts, however, the score's finite-sample distribution can be skewed or otherwise depart from that reference, especially in a tail. A small discrepancy in a tail probability matters when the evidence is judged against a stringent threshold.

Saddlepoint approximation (SPA) uses the score's cumulant-generating function to approximate its tail more accurately than a normal approximation can when the null score distribution is non-normal. It is a calibration of a test statistic's tail probability, not a new association model and not a remedy for confounding, poor phenotype definition, or a misspecified null model.

The Genetic Score and Its Null Distribution

Let \(Y_i\) be the binary case status for person \(i\), coded 1 for a case and 0 for a control. Let \(g_i\) be the genotype dosage for the allele being tested, coded 0, 1, or 2, and let \(x_i\) denote the covariate vector. Under the null hypothesis of no genetic effect, a logistic model gives a fitted case probability \(\hat{\mu}_i\) for each person. Covariate adjustment is important: a genotype association should be assessed conditional on the covariates specified in the null model.

With nuisance covariates, the score for the genetic effect is made efficient by accounting for the information shared between the genotype and those covariates. One way to express the resulting score uses residualized genotype weights. Write \(W\) for the diagonal matrix with entries \(\hat{\mu}_i(1-\hat{\mu}_i)\), and \(X\) for the covariate design matrix. Define \(\tilde{g}=g-X(X^\mathsf{T}WX)^{-1}X^\mathsf{T}Wg\), assuming the inverse exists. The efficient score can then be written \(U=\sum_i \tilde{g}_i(Y_i-\hat{\mu}_i)\). Under the fitted null model, its variance is approximately \(V=\sum_i\tilde{g}_i^2\hat{\mu}_i(1-\hat{\mu}_i)\), the efficient information for the genetic effect.

For SPA, treat the fitted null probabilities and residualized weights as fixed when forming the score's null cumulant-generating function. Under independent Bernoulli outcomes with those probabilities, it is

The corresponding derivatives are

At \(t=0\), \(K(0)=K'(0)=0\), and \(K''(0)=V\). Thus the ordinary score test uses the mean and variance at the center of the null distribution. SPA also uses how the cumulant-generating function changes away from zero, which carries information about skewness and other departures from a normal shape. Because nuisance parameters and their fitted values are estimated, this plug-in calculation is an approximation rather than an exact finite-sample distribution in every covariate-adjusted analysis.

The Saddlepoint Tail Approximation

For an observed score \(u\), first find the root \(\hat{t}\) of \(K'(\hat{t})=u\). This root selects an exponentially tilted distribution whose mean score equals the observed score. The tail calculation then uses two standardized quantities:

For a positive observed score, the Lugannani–Rice approximation to the upper-tail probability is

where \(\Phi\) and \(\phi\) are the standard normal cumulative distribution and density functions. The correction involving \(v\) and \(w\) adjusts the normal tail to reflect the score distribution's shape. For a negative observed score, the corresponding lower tail is obtained with the appropriately signed quantities. If the distribution is not symmetric, a two-sided p-value should be formed from both tail approximations, rather than assuming that one tail can simply be doubled.

Comparing Two-Sided Results and Choosing the Interpretation

A two-sided test asks about extremeness in either direction. When the null score distribution is symmetric, the two-sided tail probability can be approximated by twice the smaller one-sided tail. In the preceding hypothetical calculation, suppose the lower-tail SPA calculation is approximately equal to the upper tail. Then the two-sided SPA p-value is

The ordinary two-sided normal p-value was \(0.2386\). The calibrated value is lower by about \(0.0059\), but both are well above \(0.05\), so neither leads to rejection at that significance level. The difference illustrates that calibration can alter a p-value without changing the qualitative conclusion in a particular example. In a genome-wide analysis, even a small tail difference may matter when a result is near a decision threshold; whether it does depends on the threshold and the study's testing plan.

What SPA Does—and Does Not—Address

The central use of SPA here is to improve score-test tail calibration when the Gaussian approximation may be inadequate under severe case-control imbalance. The method is most useful when a score distribution is discrete or skewed and the analysis requires a tail probability, rather than only a rough central approximation. In applied work, the usual workflow is to fit the covariate-adjusted null model, construct the score and its null cumulant-generating function, solve for the saddlepoint, and compare the calibrated probability with the ordinary asymptotic result.

SPA does not repair a flawed null model. If covariates needed to control confounding are omitted, if the genetic or phenotype data are systematically mismeasured, or if the analysis handles relatedness incorrectly, a more accurate tail calculation does not remove those sources of error. Nor is SPA a replacement for Firth logistic regression: Firth's method addresses estimation under sparse data or separation, whereas the focus here is calibration of a score-test p-value. The methods answer related but distinct questions.

Common Mistake

Wrong reasoning: “The score test already standardizes by its variance, so its tail must be normal even with a very small case fraction.” Standardizing centers and scales the score, but it does not guarantee that its finite-sample distribution is normal. A statistic can have mean zero and variance one while still having skewness or heavier tails than a normal distribution.

Correction: Treat the normal p-value as a large-sample approximation, then consider SPA when imbalance makes tail accuracy a concern. Use the fitted null score distribution and the saddlepoint calculation, and report what was used. A calibrated p-value is still conditional on the correctness of the model and its assumptions; it is not proof that the null is true or false.