Why Sparse Carrier Counts Need Care
A structural variant may be observed in only a few people, even in a carefully designed case-control study. When carrier counts are small, a single participant can strongly affect the estimated association. If one of the four cells in a case-control table is zero, the usual odds-ratio calculation can be zero, infinite, or undefined, and ordinary logistic regression may have no finite maximum-likelihood estimate. These are not merely computational inconveniences: they change what can be estimated and how uncertainty should be assessed.
Two approaches are especially useful for sparse data. Fisher’s exact test evaluates a table by conditioning on its observed margins, giving an exact null distribution for an unadjusted comparison. Firth-penalized logistic regression modifies the likelihood to reduce small-sample bias and to yield finite coefficient estimates in many separated data sets. They address related problems, but they are not interchangeable: one is an exact conditional test for a table, while the other is a regression estimation method.
Set Up the Case-Control Table
For a binary carrier indicator, arrange the data with case status in rows and carrier status in columns. Let \(a\) be the number of carrier cases, \(b\) the number of noncarrier cases, \(c\) the number of carrier controls, and \(d\) the number of noncarrier controls. The sample odds ratio comparing cases with controls is \(ad/(bc)\), when the denominator is nonzero and the table is not otherwise degenerate. It is an association measure, not a risk ratio; in a case-control study, the sampling fractions generally do not identify disease risk in the source population.
| Carrier | Noncarrier | |
|---|---|---|
| Cases | \(a\) | \(b\) |
| Controls | \(c\) | \(d\) |
With positive counts in all four cells, a familiar large-sample standard error for the log odds ratio is \(\sqrt{1/a+1/b+1/c+1/d}\). This approximation relies on enough information in each cell for the sampling distribution to be adequately approximated by a normal distribution. When counts are very small, or a cell is zero, that approximation can be unreliable or not defined. In particular, inserting a zero into the formula does not produce a usable finite standard error.
Exact Conditional Testing
For an unadjusted \(2\times2\) table, Fisher’s exact test considers the distribution of carrier cases under the null hypothesis of no association, conditional on the observed row and column totals. Under that null and conditional on the margins, the allocation of the observed total number of carriers between cases and controls follows a hypergeometric distribution. If \(X\) is the number of carrier cases, \(n_1=a+b\) is the number of cases, \(n_0=c+d\) the number of controls, and \(m=a+c\) the total number of carriers, then
A one-sided test sums probabilities in the direction specified by the alternative hypothesis. A common two-sided Fisher test sums the probabilities of all tables with the same margins whose probability under the null is no greater than the observed table’s probability. Software can use other documented two-sided conventions, so the convention should be reported. “Exact” refers to the conditional null distribution; it does not mean that the study is free of bias or that the test automatically adjusts for covariates.
Worked Example: One Carrier Case and No Carrier Controls
Suppose a hypothetical study includes 20 cases and 20 controls. One case carries the variant, while no control does. The table is \(a=1\), \(b=19\), \(c=0\), \(d=20\). The sample odds ratio is \(1(20)/(19(0))\), which is infinite because its denominator is zero. Thus the log odds-ratio estimate is plus infinity; the odds ratio itself is not a finite estimate.
For Fisher’s test, there is one carrier in total, so \(m=1\), and there are 20 people in each group. Under the null, the single carrier is equally likely to be in either group:
For the common probability-ordering two-sided test, both possible tables have probability 0.5, so the p-value is \(0.5+0.5=1.00\). The observed table has an infinite sample odds ratio, but with only one carrier it is not unusual under this conditional null distribution. The test does not establish that there is no association; it shows that these sparse data provide little evidence against the null by this test.
Firth-Penalized Logistic Regression
For individual \(i\), let \(Y_i=1\) denote case status and \(Y_i=0\) control status. A logistic regression model specifies \(P(Y_i=1\mid x_i)=\pi_i\), with \(\operatorname{logit}(\pi_i)=\log\{\pi_i/(1-\pi_i)\}=x_i^\mathsf{T}\beta\). Here \(x_i\) is the predictor vector, which may include a carrier indicator and, when appropriate, prespecified covariates. The usual likelihood assumes independent observations conditional on the predictors and a correctly specified logistic model. In a case-control design, the slope for a predictor can estimate a log odds ratio under the usual sampling assumptions, but the intercept does not generally estimate population disease prevalence.
Ordinary logistic maximum likelihood can fail under complete separation: some linear combination of predictors perfectly distinguishes cases from controls. In a simple carrier comparison, if all carriers are cases and all noncarriers are controls, increasing the carrier coefficient can keep improving the likelihood as that coefficient tends to plus infinity. There is no finite ordinary maximum-likelihood estimate for that coefficient. Under quasi-complete separation, a boundary pattern similarly prevents an ordinary finite estimate for one or more coefficients.
The penalty discourages estimates at the boundary where fitted probabilities approach zero or one and the information becomes deficient. Consequently, Firth estimates under separation are finite and generally less extreme than the diverging ordinary estimates. This is a change in estimation, not evidence that the data contain more carriers. In an unadjusted binary-carrier model, the Firth odds ratio has a particularly simple form: add one half to each of the four table cells and calculate the odds ratio from those adjusted counts. This identity is specific to that model and should not be applied indiscriminately to more complex regression designs.
Worked Example: A Finite Firth Estimate Under Separation
Use the hypothetical table from the first example: one carrier case and no carrier controls, with 19 noncarrier cases and 20 noncarrier controls. Ordinary logistic regression has quasi-complete separation for the carrier indicator, so its maximum-likelihood carrier coefficient has no finite value. For this unadjusted binary-predictor model, the Firth odds ratio is calculated by adding 0.5 to each cell:
The corresponding carrier coefficient is \(\log(3.154)\approx1.148\). In this fitted model, the estimated odds of case status for carriers are about 3.15 times those for noncarriers. This is an association estimate, not a causal effect or a risk ratio. The finite result avoids the infinite ordinary estimate, but the study still contains only one carrier, so substantial uncertainty remains.
For uncertainty, use a method appropriate to the penalized model, such as a profile penalized-likelihood interval, and report its construction. Do not report an ordinary Wald interval based on the zero-cell standard-error formula: that formula is undefined here. A finite Firth point estimate does not by itself imply a narrow interval or strong evidence of association.
Zero Counts Can Point in Either Direction
A zero in a different cell has a different effect on the odds ratio. It can make the odds ratio zero rather than infinite. This distinction matters both for interpretation and for choosing a valid inferential method.
Worked Example: Zero Carrier Cases and a Finite Zero Odds Ratio
Consider a second hypothetical study with 30 cases and 30 controls. No case carries the variant, while four controls do. The table has \(a=0\), \(b=30\), \(c=4\), and \(d=26\). The sample odds ratio is
Zero is a finite value for the odds ratio. Its logarithm, however, is minus infinity, so there is no finite log odds-ratio estimate from this table. This example is not a degenerate table: both rows contain observations and the ratio is defined.
For a conditional exact calculation, the total number of carriers is \(m=4\), with 30 cases and 30 controls. Under the null, the probability of observing zero carrier cases is
A one-sided test for depletion of carriers among cases therefore has p-value 0.0562. For the common two-sided probability-ordering test, the table with all four carriers in the cases has the same probability, 0.05619; the other possible tables have greater probabilities. The two-sided p-value is therefore approximately \(0.05619+0.05619=0.1124\). This calculation is conditional on the observed margins and does not adjust for covariates.
Small Counts Without a Zero Cell
Sparse data can also be problematic when every cell is positive. A large-sample log-odds-ratio interval may then be computable, but its approximation can still be poor because one cell contributes little information. The next example illustrates how the ordinary estimate and Firth estimate can differ even without separation.
Worked Example: Comparing Ordinary and Firth Estimates
Suppose, hypothetically, that a study observes two carrier cases, 28 noncarrier cases, one carrier control, and 29 noncarrier controls. All four cells are positive, so the ordinary odds ratio is defined:
The ordinary log odds-ratio estimate is \(\log(2.071)\approx0.728\). Its usual large-sample standard error is \(\sqrt{1/2+1/28+1/1+1/29}=\sqrt{1.5702}\approx1.253\). The single carrier control contributes the term \(1/1\), making the estimated uncertainty large. This standard error relies on a large-sample approximation, which is questionable with these small counts.
For the unadjusted binary-predictor Firth estimate, add 0.5 to each cell:
The Firth estimate is less extreme than the ordinary estimate. Neither point estimate removes the uncertainty created by having only three carriers in total. A suitable interval and inferential method should be reported rather than interpreting the difference between the two point estimates as evidence for or against the association.
How to Choose and Report an Analysis
For a single unadjusted carrier comparison, an exact conditional test is a natural choice when counts are sparse, especially when a cell is zero. It yields a valid conditional p-value without relying on a large-sample normal approximation. It does not supply a covariate-adjusted estimate, and the p-value alone does not describe the magnitude or precision of an association.
When regression adjustment is needed, or when a finite coefficient estimate is important under separation, Firth logistic regression can be appropriate if the model is identifiable and its predictors are justified. Report the model, the Firth estimate (often as an odds ratio with units or a clear predictor contrast), and an uncertainty interval based on penalized likelihood or another defensible method. A regression model with sparse data can still be unstable, particularly if it includes many parameters relative to the number of informative observations. Penalization does not rescue an overcomplicated or nonidentifiable model.
A clear report should distinguish the analysis target and its limitations. For example, state whether the result is an unadjusted conditional test or an adjusted penalized regression estimate; give the carrier and noncarrier counts in cases and controls; state the direction and size of the odds ratio; and describe how uncertainty was calculated. The exact-test p-value and Firth estimate answer related but different questions, so do not present them as though one were a direct conversion of the other.
Common Mistakes
- Calling every zero-cell odds ratio infinite. A zero numerator can make the odds ratio zero, which is finite; its logarithm is minus infinity. A zero denominator can make the odds ratio infinite. In a degenerate table, the ratio can be undefined.
- Treating an infinite ordinary estimate as proof of an infinite biological effect. Under separation, the likelihood has no finite ordinary maximum-likelihood coefficient. The boundary estimate reflects sparse data and the model, not a known infinite association.
- Using a zero-cell Wald standard error. The usual reciprocal-cell formula is undefined when a cell is zero. Use an inference method suited to the sparse-data analysis.
- Calling a Fisher p-value the probability that the null is true. It is a tail probability under the null, conditional on the observed margins and the chosen two-sided convention.
- Assuming Firth regression fixes every problem. It addresses small-sample bias and separation under suitable model conditions; it does not correct confounding, poor variant calls, selection bias, or a nonidentifiable design.
Check Your Understanding
Try these questions without looking back.
- If \(a=0\), \(b=12\), \(c=3\), and \(d=15\), what is the sample odds ratio, and is its logarithm finite?
- In a \(2\times2\) carrier table, what is held fixed when Fisher’s exact test calculates its conditional null distribution?
- Why can ordinary logistic maximum likelihood have no finite carrier coefficient when all carriers are cases and all noncarriers are controls?
- For the simple unadjusted binary-carrier model, how is the Firth odds ratio calculated from the four table counts?
- Why does a finite Firth estimate not imply that a sparse study has little uncertainty?