Nested Case–Control Study Design

The Nested Case–Control Study Design is a two-phase epidemiological and biostatistical sampling architecture introduced by Frederick D. K. Liddell, J. Corbett McDonald, and Duncan C. Thomas in 1977 and placed on a counting-process martingale foundation by Norman E. Breslow, Bryan Langholz, Larry Goldstein, and Ørnulf Borgan. In a large prospective cohort study following N individuals over calendar or follow-up time, measuring expensive molecular biomarkers, genomic sequences, archived serum metabolites, or detailed occupational histories on all N cohort members is often cost-prohibitive when the disease of interest is rare. The nested case–control (NCC) design resolves this resource constraint by assaying expensive covariates on 100% of incident disease cases that arise during follow-up, together with a small random sample of m time-matched controls drawn from the risk set of cohort members who remain disease-free and under active observation at the observed failure time of each case. Through this incidence density sampling (or risk-set sampling) protocol, conditional logistic regression on the sampled risk sets estimates the hazard ratio parameters of the Cox proportional hazards model without rare-disease approximations and free of retrospective recall or selection bias.

Consider a well-defined underlying cohort C = {1, 2, …, N} of N individuals tracked over a follow-up time scale t ∈ [0, τ] (such as age, time since enrollment, or calendar time). For each subject j ∈ C, let T̃_j = min(T_j, U_j) denote the observed exit time, where T_j is the true event (failure) time, U_j is the right-censoring time, and δ_j = I(T_j ≤ U_j) is the failure indicator, accompanied by a vector of possibly time-varying covariates Z_j(t) = (Z_j1(t), …, Z_jp(t))^⊤.

At any time t ≥ 0, the full-cohort risk set R(t) consists of all cohort members who are under active observation and free of the outcome immediately prior to time t:

R(t) = { j ∈ C : t_0j < t ≤ T̃_j },

where t_0j denotes left-truncation (delayed entry) time. Suppose D distinct incident cases fail during the study at ordered event times t₁ < t₂ < … < t_D, with subject i_k experiencing the failure at time t_k.

Incidence Density Sampling Protocol

At each observed failure time t_k (for k = 1, …, D), the nested case–control design constructs a sampled risk set R̃(t_k) of size m + 1 as follows:

1. Include the failing index case i_k in R̃(t_k).
2. From the remaining |R(t_k)| − 1 non-failing subjects currently at risk in R(t_k) \ {i_k}, select m controls uniformly at random without replacement within the risk set R(t_k).
3. Perform the Sampling Independently Across Failure Times: The random draws at distinct event times t₁, …, t_D are conducted conditionally independently given the full-cohort failure and censoring histories.

Because sampling is independent across distinct failure times t_k, a cohort member j who survives to a late time faces multiple independent selection draws:

• Controls Later Becoming Cases: A subject selected as a control in R̃(t_k) at an early failure time t_k remains eligible to become an incident case at a later failure time t_r > t_k if they subsequently develop the disease.
• Repeated Control Selection: A disease-free subject may be sampled as a control in more than one sampled risk set R̃(t_k) and R̃(t_r). Excluding future cases from control eligibility at time t_k violates the martingale risk-set structure and biases the estimator away from the true hazard ratio toward an cumulative odds ratio.

Suppose the individual hazard rate λ_j(t) for subject j ∈ C follows the Cox proportional hazards model:

λ_j(t | Z_j(t)) = λ₀(t) exp(β^⊤ Z_j(t)),

where λ₀(t) is an unspecified baseline hazard function and β ∈ ℝ^p is the vector of log-hazard ratios. In the full cohort, Cox's partial likelihood compares each failing case i_k against all subjects in R(t_k).

As established by Duncan C. Thomas (1977), Larry Goldstein and Bryan Langholz (1992), and Ørnulf Borgan, Goldstein, and Langholz (1995) using marked point processes, when each sampled risk set R̃(t_k) of size m + 1 is constructed by simple random sampling of m controls from R(t_k) \ {i_k}, the conditional probability that subject i_k is the failing case at time t_k given the composition of R̃(t_k) and the history up to t_k⁻ simplifies because the uniform sampling probabilities cancel between numerator and denominator. Multiplying these conditional probabilities across all D failure times yields the nested case–control partial likelihood:

L_NCC(β) = ∏_k=1^D [ exp(β^⊤ Z_ik(t_k)) / ∑_{j ∈ R̃(tk)} exp(β^⊤ Z_j(t_k)) ].

This function L_NCC(β) is algebraically identical to the conditional logistic regression likelihood for a 1 : m matched case–control study in which each sampled risk set R̃(t_k) forms a matched stratum of 1 case and m controls. Crucially, exp(β_r) estimates the instantaneous hazard ratio (incidence rate ratio) rather than a prevalence odds ratio, with no requirement that the cumulative disease incidence be small.

Counter-Matching and Stratified Risk-Set Sampling

Langholz and Borgan (1995) generalized simple random risk-set sampling to counter-matched nested case–control designs. When an inexpensive proxy variable or exposure surrogate X_j (such as job title or registry radiation dose) is available for all N cohort members in Phase 1, the risk set R(t_k) can be partitioned into L exposure strata R₁(t_k), …, R_L(t_k). If the case i_k falls in stratum s, the investigator samples m_ℓ controls from each stratum ℓ so that every stratum contains a fixed total number of subjects in R̃(t_k). Introducing known stratum inclusion weights w_j(t_k) = |R_ℓ(t_k)| / m_ℓ into the denominator of L_NCC(β) yields a weighted partial likelihood that substantially reduces variance for rare exposures and gene–environment interactions.

Let β̂_NCC denote the maximum partial likelihood estimator maximizing L_NCC(β), and let β̂_Full denote the full-cohort Cox estimator. Under regularity conditions proved by Goldstein and Langholz (1992), √D (β̂_NCC − β) converges in distribution to a multivariate normal law N(0, Σ_NCC(m)).

At the null hypothesis β = 0 for a single scalar covariate, the asymptotic relative efficiency (ARE)—defined as the ratio of the asymptotic variance of the full-cohort estimator to that of the nested case–control estimator with m controls per case—follows the famous Ury–Breslow efficiency law:

ARE(β̂_NCC vs. β̂_Full | β = 0) = m / (m + 1).

With m = 4 controls per case, a nested case–control study achieves 80% of the statistical efficiency of assaying the entire cohort while reducing biospecimen and laboratory costs by more than 90% when the outcome occurs in fewer than 2% of cohort members. When |β| is large or when analyzing rare binary exposures and multiplicative interactions, selecting m = 5 to m = 10 controls per case (or employing counter-matching) preserves high efficiency away from the null.

In its classical conditional logistic formulation, each control sampled in R̃(t_k) is tied to the specific failure time t_k of its matched case, which historically limited nested case–control studies to estimating relative hazards for a single disease outcome. Two methodological advances removed this limitation:

1. Langholz–Borgan Baseline Cumulative Hazard Estimator (1997): To estimate the absolute cumulative baseline hazard Λ₀(t) = ∫₀^t λ₀(u) du and absolute survival curves from a nested case–control sample, Langholz and Borgan introduced an up-weighted Breslow estimator:

Λ̂₀(t) = ∑_{tk ≤ t} [ 1 / ( w(t_k) ∑_{j ∈ R̃(tk)} exp(β̂^⊤ Z_j(t_k)) ) ],

where w(t_k) = |R(t_k)| / (m + 1) scales the sampled risk set of size m + 1 back up to the cardinality |R(t_k)| of the full-cohort risk set at time t_k.

1. Samuelsen's Inclusion Probabilities (1997): Sven Ove Samuelsen observed that the union of all subjects ever sampled as controls across the D risk sets forms a well-defined probability sample of the non-cases in the cohort. Because risk-set sampling is independent across failure times t₁, …, t_D, the marginal probability p_j that a non-case subject j ∈ C is selected as a control in at least one sampled risk set R̃(t_k) is:

p_j = 1 − ∏_{tk : j ∈ R(tk) \ {ik}} ( 1 − m / (|R(t_k)| − 1) ),

with p_j = 1 for all incident cases. Weighting each ever-sampled subject j by the Horvitz–Thompson inverse probability weight w_j = 1 / p_j breaks the original case-control matching ties and allows all pooled NCC controls to be reused to analyze secondary disease endpoints, fit accelerated failure time models, and construct Kaplan–Meier absolute risk curves.

The nested case–control design and the Prentice weighted case–cohort design (1986) are the two flagship outcome-dependent sampling architectures in modern molecular epidemiology:

Because nested case–control matching pairs each case with controls who have survived to the identical follow-up duration and permits simultaneous matching on blood draw date, fasting status, and freezer storage time, the NCC design is particularly advantageous in metabolomic, proteomic, and serological studies where pre-analytical sample degradation or batch drift over time could otherwise confound biomarker associations.

1. Liddell, F. D. K., McDonald, J. C., & Thomas, D. C. (1977). Methods of cohort analysis: Appraisal by application to asbestos mining. Journal of the Royal Statistical Society: Series A (General), 140(4), 469–491. https://doi.org/10.2307/2345280
2. Goldstein, L., & Langholz, B. (1992). Asymptotic theory for nested case-control sampling in the Cox regression model. The Annals of Statistics, 20(4), 1903–1928. https://doi.org/10.1214/aos/1176348895
3. Borgan, Ø., Goldstein, L., & Langholz, B. (1995). Methods for the analysis of sampled cohort data in the Cox proportional hazards model. The Annals of Statistics, 23(5), 1749–1778. https://doi.org/10.1214/aos/1176324322
4. Langholz, B., & Borgan, Ø. (1995). Counter-matching: A stratified nested case-control sampling method. Biometrika, 82(1), 69–79. https://doi.org/10.1093/biomet/82.1.69
5. Samuelsen, S. O. (1997). A pseudolikelihood approach to analysis of nested case-control studies. Biometrika, 84(2), 379–394. https://doi.org/10.1093/biomet/84.2.379