Case-Control Imbalance and Effective Sample Size

Why Imbalance Matters

In a case-control genome-wide association study, the total number of participants does not by itself describe how much information the study contains. A study with many more controls than cases can have less association precision than a study of the same total size with a more balanced allocation. This matters when planning recruitment: the practical question is not only how many people can be enrolled, but which group an additional participant should join.

The effective sample size gives a convenient approximation for making this comparison. It converts unequal case and control counts into the size of a hypothetical balanced study with similar information for a genetic association. It is a measure of precision under specified assumptions, not a replacement for the actual numbers enrolled and not a claim that different designs are identical in every respect.

Definition: For $n_{\text{case}}$ cases and $n_{\text{control}}$ controls, the case-control effective sample size is $$ n_{\text{eff}}=\frac{4n_{\text{case}}n_{\text{control}}}{n_{\text{case}}+n_{\text{control}}} =\frac{4}{1/n_{\text{case}}+1/n_{\text{control}}}. $$ It is the total sample size of a balanced case-control study that has approximately the same association information under the assumptions described below.

Deriving the Approximation

Consider an additive genetic association analysis using a genotype dosage $G$, coded as 0, 1, or 2 copies of an effect allele. Suppose cases and controls are independent samples, and under the null hypothesis the dosage has the same variance, $\sigma_G^2$, in the two groups. The difference between the case and control mean dosages has variance

\operatorname{Var}(\bar{G}_{\text{case}}-\bar{G}_{\text{control}}) =\sigma_G^2\left(\frac{1}{n_{\text{case}}}+\frac{1}{n_{\text{control}}}\right).

This follows because the variance of a sample mean from $n$ independent observations is $\sigma_G^2/n$, and the variance of the difference between independent sample means is the sum of their variances. For a balanced study with $m$ cases and $m$ controls, the total sample size is $N=2m$, so the same variance is $4\sigma_G^2/N$. Equating the two variance expressions identifies the balanced total size that gives the same variance:

\sigma_G^2\left(\frac{1}{n_{\text{case}}}+\frac{1}{n_{\text{control}}}\right) =\frac{4\sigma_G^2}{n_{\text{eff}}} \quad\Longrightarrow\quad n_{\text{eff}}=\frac{4}{1/n_{\text{case}}+1/n_{\text{control}}} =\frac{4n_{\text{case}}n_{\text{control}}}{n_{\text{case}}+n_{\text{control}}}.

The same allocation factor appears in the usual local information approximation for a logistic association model. Under the null, with an intercept and a genetic dosage as predictor, the information for the dosage coefficient after accounting for the intercept is approximately $\sigma_G^2 n_{\text{case}}n_{\text{control}}/(n_{\text{case}}+n_{\text{control}})$. For a balanced study of total size $N$, that factor is $\sigma_G^2N/4$. Equating them again gives $N=n_{\text{eff}}$. Thus, under the approximation, association information is proportional to $n_{\text{eff}}$, and the standard error is proportional to $1/\sqrt{n_{\text{eff}}}$, when the other features of the analysis are held fixed.

The derivation relies on comparable genotype variance across cases and controls under the null, independent participants, and a common genetic effect and analysis model. For example, under Hardy–Weinberg equilibrium with effect-allele frequency $f$, genotype dosage variance is $2f(1-f)$. That variance factor affects the information available at a variant, but it cancels when comparing case-control allocations at the same variant. The effective-size formula isolates the contribution of the allocation itself.

Worked Example: Converting an Imbalanced Study to Effective Size

Suppose a hypothetical GWAS recruits 900 cases and 2,700 controls. Find its effective sample size and compare it with a balanced study having the same effective size.

Calculate the effective size. Substitute the two group counts:

n_{\text{eff}}= \frac{4(900)(2700)}{900+2700} =\frac{9{,}720{,}000}{3{,}600} =2{,}700.

So this design has approximately the same allocation-related association information as a balanced study with 2,700 participants in total: 1,350 cases and 1,350 controls. The actual study still has 3,600 participants. The value 2,700 is an information-equivalent balanced total, not a count of the people who were enrolled.

How Imbalance Reduces Effective Size

Let $N=n_{\text{case}}+n_{\text{control}}$ be the actual total. Define the case-to-control ratio $r=n_{\text{case}}/n_{\text{control}}$. Rearranging the formula gives

n_{\text{eff}}=\frac{4r}{(1+r)^2}N.

For a balanced allocation, $r=1$, and $n_{\text{eff}}=N$. For any other positive ratio, $4r/(1+r)^2$ is less than 1, so the effective size is smaller than the actual total. The formula is symmetric: swapping the case and control counts leaves $n_{\text{eff}}$ unchanged. It depends on how imbalanced the allocation is, not on which group is larger.

The penalty grows as the allocation becomes more unequal. For a fixed total, a balanced allocation maximizes $n_{\text{eff}}$: the product $n_{\text{case}}n_{\text{control}}$ is largest when the two counts are equal. This does not mean that a balanced design is always feasible or optimal overall. Cases may be costly or difficult to recruit, and study goals may impose other constraints. The formula quantifies the association-information consequence of the allocation, not every consideration in a recruitment decision.

Worked Example: Comparing Two Allocations at the Same Total

Two hypothetical studies each recruit 10,000 participants. Study A has 5,000 cases and 5,000 controls. Study B has 2,000 cases and 8,000 controls. Compare their effective sizes and approximate standard errors, assuming the same variant, phenotype definition, and analysis.

Study A.

n_{\text{eff,A}}= \frac{4(5000)(5000)}{10{,}000} =10{,}000.

Study B.

n_{\text{eff,B}}= \frac{4(2000)(8000)}{10{,}000} =6{,}400.

With the other information factors held fixed, standard errors scale approximately as $1/\sqrt{n_{\text{eff}}}$. Thus the ratio of Study B's standard error to Study A's is

\frac{SE_{\text{B}}}{SE_{\text{A}}} \approx \sqrt{\frac{10{,}000}{6{,}400}} =1.25.

Study B's standard error is approximately 25% larger under this approximation. The comparison is about precision, not about a 25% change in effect size or a guarantee of a particular difference in power. Exact power also depends on the effect, allele frequency, testing threshold, and other design features.

Choosing Whether to Add Cases or Controls

The effective-size formula makes the marginal value of recruitment visible. If one group is much smaller, increasing that group generally raises effective size more than adding the same number to the already larger group. However, additional participants in either group can increase information; the gain is simply not equal for all allocations. The formula can compare alternatives before committing resources.

Worked Example: Adding Cases or Adding Controls

A hypothetical study starts with 1,000 cases and 9,000 controls. Investigators can recruit either 1,000 additional cases or 9,000 additional controls. Compare the effective-size gains for these two options.

Starting design.

n_{\text{eff,start}}= \frac{4(1000)(9000)}{1000+9000} =\frac{36{,}000{,}000}{10{,}000} =3{,}600.

Option 1: add 1,000 cases. The new counts are 2,000 cases and 9,000 controls:

n_{\text{eff,1}}= \frac{4(2000)(9000)}{2000+9000} =\frac{72{,}000{,}000}{11{,}000} \approx6{,}545.

The gain is approximately $6{,}545-3{,}600=2{,}945$ effective participants.

Option 2: add 9,000 controls. The new counts are 1,000 cases and 18,000 controls:

n_{\text{eff,2}}= \frac{4(1000)(18{,}000)}{1000+18{,}000} =\frac{72{,}000{,}000}{19{,}000} \approx3{,}789.

The gain is approximately $3{,}789-3{,}600=189$ effective participants. In this example, even though the second option adds many more people, it yields a much smaller increase in effective size. The result does not establish which option is cheaper or more practical; it isolates the allocation's contribution to association information.

What the Effective Size Does and Does Not Say

The formula is useful for a first comparison of recruitment plans, but its interpretation is deliberately limited. Under a common-effect, small-effect approximation, studies with the same $n_{\text{eff}}$ have similar information from the case-control allocation for a given variant. This supports comparisons of approximate standard errors and power when the genetic variance, phenotype, analysis, and testing threshold are otherwise the same. It does not say that the datasets are interchangeable, or that their estimates will be identical.

The approximation can be less representative when other features differ materially. Examples include differences in ancestry composition that change allele frequencies, relatedness that reduces the amount of independent information, different phenotype accuracy, or different covariate distributions. If genotype variance differs between cases and controls, the simple common-variance derivation does not capture that difference. More detailed models may be needed when such features are central to the design.

Effective sample size is also not a direct measure of population representativeness. Case-control studies can deliberately sample cases and controls in proportions that differ from disease prevalence. That sampling choice does not, by itself, make the effective-size calculation invalid for comparing association precision, but it does mean that $n_{\text{case}}/N$ should not be read as population disease prevalence.

A useful limiting case clarifies why the smaller group matters so much. If the number of controls grows without bound while the number of cases stays fixed, then

\lim_{n_{\text{control}}\to\infty} \frac{4n_{\text{case}}n_{\text{control}}}{n_{\text{case}}+n_{\text{control}}} =4n_{\text{case}}.

So adding controls cannot make the effective size grow without limit when the case count is fixed. Under this approximation, it approaches a ceiling of four times the number of cases. This is not a rule that controls cease to matter at a particular count; it expresses diminishing returns when controls greatly outnumber cases.

Common Mistake

Wrong reasoning: “The study has 20,000 participants, so its effective sample size is 20,000.” This ignores how the participants are divided between cases and controls.
Correction: Calculate $n_{\text{eff}}=4n_{\text{case}}n_{\text{control}}/(n_{\text{case}}+n_{\text{control}})$. The effective size equals the actual total only for a balanced allocation under this approximation.
Wrong reasoning: “Adding controls always helps as much as adding cases.” The gain depends on the current allocation and the number added.
Correction: Recalculate effective size for each proposed allocation. In a highly imbalanced study, adding to the smaller group often yields a larger gain per participant.

Key takeaway: For unequal case and control counts, use $n_{\text{eff}}=4n_{\text{case}}n_{\text{control}}/(n_{\text{case}}+n_{\text{control}})$ to approximate the balanced-study size with similar allocation-related association information. Treat it as a design comparison tool, with assumptions, rather than as the study's actual sample count.

Check Your Understanding

Answer each question using the effective-size approximation and its stated interpretation.

A hypothetical study has 800 cases and 3,200 controls. What is its effective sample size?
Why does a balanced allocation maximize effective size for a fixed total sample size?
Two studies have the same effective size, but one has a different ancestry composition. Why might the effective size alone not imply equal association precision?
If the number of cases is fixed and the number of controls becomes very large, toward what value does effective size approach?
A researcher says that a case-control study's case fraction estimates disease prevalence. Explain why this conclusion does not follow from the effective-size calculation.