Firth Logistic Burden Tests Under Separation

Why Burden Tests Can Need Penalization

A burden test combines information across variants into a subject-level predictor. This can be useful when a set of rare alleles is expected to have effects in a common direction, but the resulting predictor may be carried by very few people. If all burden carriers in a sample are cases, or all are controls, ordinary logistic maximum likelihood can fail: the estimated burden coefficient may grow without bound rather than settle at a finite value. Strong case-control imbalance and sparse carrier patterns can also make ordinary estimates highly variable.

Firth logistic regression modifies the likelihood with a penalty that reduces first-order estimation bias. In logistic models, it is particularly useful because the penalized estimate is often finite even when the ordinary maximum-likelihood estimate is not. It does not, however, create information that the data do not contain, correct a misspecified burden, or remove confounding.

The burden’s unit depends on its construction. With unweighted allele dosages, a one-unit increase means one additional counted allele. With other weights or a normalized burden, it means one unit on that chosen scale. The model also imposes a substantive restriction: the variants contribute through this single aggregate predictor, so variant-specific directions and effects are not separately estimated by this burden coefficient.

Ordinary and Firth Logistic Likelihoods

Write \(x_i^\mathsf{T}\) for row \(i\) of the design matrix, including the intercept, burden, and covariates; write \(\boldsymbol{\beta}\) for their coefficients; and let \(\mu_i=\operatorname{logit}^{-1}(x_i^\mathsf{T}\boldsymbol{\beta})\). For independent binary outcomes with the stated logistic mean model, the ordinary log-likelihood is

Firth estimation maximizes the penalized log-likelihood

Here \(I(\boldsymbol{\beta})\) is the model’s expected information matrix. The penalty is one-half the log determinant of that matrix, also called the Jeffreys-prior penalty. This is a penalized-likelihood estimation procedure, not a claim that a particular prior distribution has been imposed on the burden coefficient.

Why the Score Is Adjusted

The ordinary score is \(U(\boldsymbol{\beta})=X^\mathsf{T}(Y-\boldsymbol{\mu})\). Differentiating the penalty adds a leverage-dependent adjustment. Let \(h_i\) be the \(i\)th diagonal element of the weighted hat matrix \(W^{1/2}X(X^\mathsf{T}WX)^{-1}X^\mathsf{T}W^{1/2}\). The Firth score is

where \(\circ\) denotes elementwise multiplication. This result follows by differentiating \(\frac12\log|I|\): the matrix identity \(d\log|I|=\operatorname{tr}(I^{-1}dI)\), applied to \(I=X^\mathsf{T}WX\), gives a contribution \(x_i h_i(1/2-\mu_i)\) for observation \(i\). Summing those contributions gives the adjustment displayed above. The term is largest in relevance for observations with high leverage; it modifies the score rather than simply adding a constant to every coefficient.

Under complete or quasi-complete separation, the ordinary likelihood can keep increasing as a coefficient moves toward an infinite limit. The information and fitted probabilities then become degenerate. The Firth penalty changes the estimating equation and commonly yields finite estimates in full-rank logistic designs with both outcome classes represented. A rank-deficient design, an outcome with no cases or no controls, or other severe data pathologies still require attention; penalization is not a universal guarantee of an estimable scientific contrast.

Testing the Burden Coefficient

Fit a reduced model containing the intercept and covariates but not the burden, and a full model that adds \(B_i\). In both fits, use the Firth-penalized likelihood. The penalized likelihood-ratio statistic is

Because the reduced and full fits use different-dimensional information matrices and hence different Jeffreys penalties, the difference of separately maximized penalized log-likelihoods does not in general have the ordinary Wilks chi-square reference distribution; the penalty difference can be of order \(\log n\) even under \(H_0\). Use a testing procedure with justified calibration, such as a parametric bootstrap under the null that refits both models in each replicate, or a consistently defined profile penalized-likelihood procedure. A Wald test instead divides the estimated coefficient by its standard error and uses a normal reference; it can be especially fragile when the coefficient is large or the likelihood is asymmetric. Small samples, very sparse carriers, or extreme imbalance can make either approximation unreliable.

An odds ratio is conditional on included covariates and on the burden scale. A test p-value is the probability, under the null model and the test’s assumptions, of observing a statistic at least as extreme as the one calculated. It is not the probability that the null hypothesis is true. Statistical significance also does not establish that the aggregate burden is causal or that every variant in the set has the same effect.

Assumptions, Diagnostics, and Alternatives

Firth penalization changes estimation, not the basic modeling assumptions. For the usual logistic interpretation and test, subjects should be independent conditional on modeled structure; the outcome should be binary; the logit mean model should be adequately specified; the burden and covariates should be defined consistently; and the design must identify the coefficient being tested. Related subjects, repeated measurements, or other dependence require a model that addresses that dependence. Important confounders and appropriate population-structure covariates still need to be considered.

Before interpreting a result, inspect the burden distribution by outcome and the number of carriers and noncarriers among cases and controls. Look for empty or nearly empty cells, a burden concentrated in a few subjects, strong collinearity with covariates, and coefficients or standard errors that change sharply under reasonable model checks. Confirm that the design matrix is full rank and that both outcome classes occur. A finite Firth estimate does not itself show that the burden is well measured, that the logit relationship is suitable, or that the selected variant set and weights are scientifically justified.

The main alternative is ordinary maximum-likelihood logistic regression. When data are well behaved, it is a standard approach and avoids the Firth penalty; under separation, it may have no finite estimate, and in sparse settings its estimates and Wald inference can be unstable. Firth logistic regression offers bias reduction and often finite estimates, but the penalty can materially affect estimates in small samples, and approximate p-values may still be poorly calibrated. Exact logistic methods are another option for some small, sparse settings, but can be computationally demanding and may condition away information. Neither alternative substitutes for choosing a burden model that matches the scientific question.

Putting the Analysis in Context

A Firth burden analysis is most defensible when the burden definition and weights have a clear rationale, the logistic covariate model is appropriate, and sparse data make ordinary maximum likelihood unreliable. Report the burden scale, the fitted effect and its uncertainty, the test used, and the limitations created by sparse carriers or case-control imbalance. A burden coefficient summarizes association with the aggregate predictor; it does not reveal which variant drives the signal or prove a shared effect direction across variants.