Bias in Meta‐Analytic Modeling of Surrogate Endpoints in Cancer Screening Trials Authors: James P Long, Abhishikta Roy, Ehsan Irajizad, Kim‐Anh Do, Yu Shen Published: 2026 ABSTRACT ABSTRACTIn meta‐analytic modeling, the functional relationship between a primary and surrogate endpoint is estimated using summary data from a set of completed clinical trials. Parameters in the meta‐analytic model are used to assess the quality of the proposed surrogate. Recently, meta‐analytic models have been employed to evaluate whether late‐stage cancer incidence can serve as a surrogate for cancer mortality in cancer screening trials. A major challenge in meta‐analytic models is that uncertainty of trial‐level estimates affects the evaluation of surrogacy, since each trial provides only estimates of the primary and surrogate endpoints rather than their true parameter values. In this work, we show via simulation and theory that trial‐level estimate uncertainty may bias the results of meta‐analytic models toward positive findings of the quality of the surrogate. We focus on cancer screening trials and the late‐stage incidence surrogate. We reassess correlations between primary and surrogate endpoints in Ovarian cancer screening trials. Our findings indicate that completed trials provide limited information regarding quality of the late‐stage incidence surrogate. These results support restricting meta‐analytic regression usage to settings where trial‐level estimate uncertainty is incorporated into the model. Introduction Cancer screening aims to reduce mortality by detecting tumors at an early stage, when treatments are more likely to be effective. Screening programs have been widely used in clinical practice including mammograms for breast cancer, sigmoidoscopy for colorectal cancer, and low‐dose computed tomography (LDCT) scans for lung cancer in smokers [1]. New imaging and molecular diagnostic technologies have led to the development of promising new screening techniques, which are at various stages of scientific and commercial development [2, 3, 4]. Before being widely adopted, a screening method should be tested in clinical trials to demonstrate benefit to people at average or increased risk of the cancer. These trials typically randomize subjects to two arms: a standard of care (control) arm and a screening (intervention) arm. Cancer‐specific mortality is often considered the gold‐standard primary endpoint in screening trials [5, 6]. Unfortunately, using reduction in cancer‐specific mortality to demonstrate efficacy is costly, both in length of follow‐up time and trial sample size. For example, the National Lung Screening Trial (NLST) demonstrated a 20.0% decrease in lung cancer‐specific mortality in the screen arm (annual LDCT) compared with the control arm (annual chest radiography) [7]. The trial enrolled over 50,000 healthy subjects with more than 7 years of follow‐up. In another example, the Health Insurance Plan (HIP) mammography trial enrolled 62,000 women with 18 years of follow‐up [8]. Surrogate (also known as alternative) endpoints, which can be determined at the time of cancer diagnosis (e.g., tumor stage or size), may facilitate faster evaluation of efficacy than mortality endpoints. Among possible surrogates, late‐stage cancer incidence has emerged as perhaps the most promising choice [9, 10, 11, 12, 13, 14]. Use of this surrogate is motivated by the fact that tumors diagnosed at early stage have better prognosis than tumors diagnosed at late stage. For example, ovarian cancer diagnosed at early‐stage has approximately a 90% 5‐year survival probability while ovarian cancer diagnosed at a late‐stage has approximately a 30% 5‐year survival probability [15]. However, there is no guarantee that a screening modality which reduces late‐stage incidence will also reduce mortality. For example, [16] notes that certain biomarkers are only released by aggressive tumors and that these cancers, even if detected early, may remain deadly. A liquid biopsy screening test which detects these biomarkers may reduce late‐stage incidence but not affect mortality. In this case, late‐stage reduction is a poor surrogate which could produce a misleading positive result. This concern has prompted systematic investigation of the suitability of late‐stage incidence as a surrogate for cancer‐specific mortality, including meta‐analyses of randomized screening trials [9, 10, 12]. These works have used a technique known as meta‐analytic modeling [17, 18] to validate the late‐stage incidence surrogate. In these meta‐analytic models, estimated mortality reduction was shown to correlate with estimated late‐stage incidence reduction in screening trials. Quantities such as regression slope and significance of positive association between primary and surrogate are reported. These regression analyses are often performed by cancer type (Ovarian, Breast, etc.), to evaluate possible cancer type specific differences in surrogate‐primary correlations. For example, [12] found significant correlation between estimated mortality reduction and estimated late‐stage incidence reduction in Ovarian screening trials (slope 1.16, 95% CI on Pearson correlation (0.51,1.00)) but not in colorectal screening trials (slope 0.4, 95% CI on Pearson correlation (−0.27,0.80)). Based on these results, [12] concluded that late‐stage incidence may be a suitable alternative endpoint for ovarian but not for colorectal cancer screening. For each trial, the estimated mortality reduction and estimated late‐stage incidence reduction generally do not exactly equal the true mortality reduction and true late‐stage incidence reduction of the screening method, due to trial‐level estimate uncertainty. This uncertainty will be large in trials which enroll few subjects and/or with low cancer incidence. In this work, we demonstrate how trial‐level estimate uncertainty affects inferences derived from meta‐analytic models, with a focus on the practice of regressing estimated mortality reduction on estimated late‐stage incidence reduction. We show that trial‐level estimate uncertainty could bias regression slopes in a positive direction under the null hypothesis, potentially leading to a false indication of surrogacy. This positive bias is demonstrated in simulations and proved analytically. We then derive findings on the use of late‐stage incidence surrogate in ovarian cancer, using confidence regions to assess the impact of uncertainty. This work is organized as follows. In Section 2, we review meta‐analysis of surrogate and primary endpoints. In Section 3, we introduce notation. In Section 4, we show via simulation that meta‐analysis which ignores trial‐level estimate uncertainty can lead to erroneous conclusions. In Section 5, we prove the generality of our simulation results under weak conditions. In Section 6, we construct 2‐D confidence regions for the surrogate–primary pair and reassess the evidence for use of late‐stage incidence as a surrogate endpoint in ovarian cancer. We conclude with a discussion in Section 7. Meta‐Analytic Modeling of Surrogate Endpoints and Trial‐Level Estimate Uncertainty Surrogate endpoints have been considered in clinical trials of treatments for many chronic diseases including AIDS, heart disease, and cancer [19, 20, 21, 22, 23, 24, 25, 26]. Before use in practice, it is critical to evaluate the quality of a candidate surrogate endpoint and its relationship to the primary endpoint. Joffe and Greene [17] summarized four approaches to evaluating surrogate quality, including meta‐analytic modeling of completed studies. VanderWeele [18] argued that among the four approaches, meta‐analytic modeling may be the most promising. Meta‐analytic modeling of surrogates is complex and differs substantially from conventional meta‐analysis of treatment efficacy. In meta‐analysis of treatment efficacy, the objective is to pool information across trials to obtain an overall estimate of an efficacy parameter. For example, let M^1,…,M^nT be estimated mortality reductions due to screening across nT trials. Under the assumption of homogeneity, all screening methods share the same cancer mortality reduction M. The estimates M^1,…,M^nT may be combined (e.g., using weighted average) to construct an estimate of M which is more efficient than any individual estimate M^l. More sophisticated random effects models may be employed in the case of heterogeneity, when the true efficacy M varies across trials [27, 28]. The purpose of meta‐analytic modeling of surrogates is to estimate the functional relationship between a surrogate and primary endpoint across trials [17, 18, 20, 24, 29]. Let Ml denote the mortality reduction for trial l. For example, if trial 1 is annual mammogram screening (with control arm receiving standard of care) while trial 2 is biannual mammograms (with control arm receiving standard of care), one may expect that M10. See Appendix 8.3 for a proof. This result shows that the positive regression slopes observed in Simulations A and B (Figure 2) are not particular to the parameters selected in Scenarios 1 and 2 of Table 1, but rather a general property that holds under any plausible setting: the mortality rate for late‐stage diagnoses is higher than that for early‐stage diagnoses both control and screen arms. For more complex settings where there is variation in Sl and Ml across trials (e.g., Simulations C and D), the fact that ρl>0 for each trial will bias estimates of β1. The degree of this bias is determined by the variation in Sl and Ml across trials and the size of ∑SM,l for each trial. Meta‐Analytic Regression in Ovarian Screening Trials To better assess how trial‐level estimate uncertainty could affect conclusions in the meta‐analytic regression, we propose visualizing uncertainty on estimates of primary and surrogate endpoints in screening trials. Construction of confidence sets for (S^l,M^l) requires estimates of the three scalar parameters comprising ∑SM,l (∑S,l, ∑M,l, and ρl) from Equation (6). These are functions of model parameters defined in Section 3 and characterize the sampling distribution of the estimators for each trial. (See Appendix Section 8.4 for mathematical details.) Diagonal elements (∑S,l and ∑M,l) are estimable from the marginal distribution of death and stage at diagnosis (by arm), information which is generally published in manuscripts and compiled for meta‐analyses. However, ρl, which we term the sampling distribution correlation, depends on the conditional probability of death given stage of diagnosis in the screen and control arms of trial l. Estimates of these conditional probabilities is generally not provided in publications reporting the results of screening trials. Thus, these conditional probabilities are not available for meta‐analyses. To address this limitation we conduct a sensitivity analysis, constructing confidence sets using various plausible values of the sampling distribution correlation ρl. As shown in Theorem 2, ρl>0 under very general conditions. We select values ρl=0.1, 0.66, and 0.9 for visualization with 0.1 and 0.9 representing extreme lower and upper bounds. The intermediate value of 0.66 is used because in a subject level analysis of the National Lung Screening Trial (NLST), ρ^=0.66 [7]. We plot results assuming all nT trials share the same ρl. In practice, ρl will vary across trials. Feng et al. [12] regressed M^l on S^l in four screening trials for Ovarian cancer. Pearson correlation across four screening trials was estimated to be R=0.99 (95% CI, 0.51 to 1.00) with linear regression slope of β^1=1.15 (Figure 1, Panel D from [10]). Figure 3 displays a scatterplot using the estimated (S^l,M^l) from the four ovarian cancer screening trials, adding 90% confidence regions on the trial parameter estimates with ρl=0.1 (left plot), ρl=0.66 (center plot), and ρl=0.9 (right plot). PLCO and UKCTOCS (USS) confidence regions include (0,0) at all correlations considered. UKCTOCS (MMS) confidence region contains M=0 at all correlations considered. The coverage properties of the UKOCST (CA‐125) confidence region may not equal the nominal level of 90% due to the small sample size of this trial (18 cancer deaths in control, 9 cancer deaths in screen). The size of the confidence regions shows that trial‐level estimate uncertainty could lead to a false but significantly positive slope found in [10]. A seemingly positive association between trial‐level late‐stage incidence reduction (as a surrogate) and mortality reduction may not guarantee a useful surrogate for the mortality reduction, as demonstrated in the ovarian cancer screening trials. Estimated mortality reduction versus estimated late‐stage incidence reduction for ovarian screening trials with 90% confidence sets for each trial estimates. Three values of ρl are considered, assumed common across all four trials. Discussion It is well established from cancer registry databases that stage of cancer diagnosis and cancer lethality are highly correlated: cancers diagnosed at an early stage are more likely to be treated successfully, improving the overall prognosis. This fact has motivated research and development in cancer screening programs. This fact also motivates investigation of late‐stage incidence as a surrogate for reduction in mortality caused by screening. However, a strong correlation between two variables does not always imply that an intervention which alters one will alter the other. As an extreme counter‐example, while ventricular arrhythmia and mortality are strongly correlated, drugs which have successfully reduced ventricular arrhythmia have been found to increase mortality [35]. To avoid this so called “surrogate paradox” [18], meta‐analytic modelings of cancer screening trials have sought to answer the questions: If a screening program reduces late stage incidence, will it also reduce mortality? If yes, by how much? We showed via simulation and theory that trial‐level estimate uncertainty could induce misleading positive slopes in meta‐analytic regression models, even under the null setting of no reduction in both surrogate and primary outcomes across trials (Sl=Ml=0 for every trial). Existing work including [12] has argued that the positive slopes in the meta‐analysis regression provide affirmative answers to the aforementioned questions and thus support use of the proposed surrogate as an endpoint for future trials. This work calls for cautious interpretation of these claims. Hierarchical models could in principle be used to estimate regression parameters while accounting for trial‐level estimate uncertainty [20, 24, 31, 32, 36]. However, there are several challenges to implementation in the context of cancer screening trials, including lack of access to subject level study data and limited number of studies in many cancer types. In related work, [37] compared performance of meta‐analytic models for predicting treatment effect on the primary endpoint given an estimate of treatment effect on the surrogate endpoint. Their focus is somewhat different from this work, which studies bias in estimates of parameters in the meta‐analytic model itself rather than the treatment effect predictions made using the model. Bujkiewicz et al. [37] tested a variety of models including the linear model studied here Equation (3) and Bayesian hierarchical models [20]. They found that the linear model, which ignores trial‐level estimate uncertainty, produced optimistic (excessively tight) uncertainty estimates on predicted treatment effect on the primary endpoint. They concluded that meta‐analytic models should incorporate trial‐level estimate uncertainty. Our work further supports this conclusion and extends the results to cancer screening surrogate endpoints while providing theoretical support for our conclusions. This work was limited to studying the effect of trial‐level estimate uncertainty on regression slopes in meta‐analyses of cancer screening trials. We did not consider several additional challenges when evaluating suitability of the late‐stage incidence surrogate. First, different functional forms between mortality reduction and late‐stage incidence reduction may hold for different screening modalities. For example, colonoscopies and liquid biopsy screening for colorectal cancer may exhibit different relationships between primary and surrogate endpoints. There is no guarantee that either of these relationships will be linear or that there would even be a single linear relationship for all types of colonoscopy or all types of liquid biopsy. Multi‐state models (MSM) which simulate patient level outcomes could potentially be used to model the primary‐surrogate relationship [38, 39, 40]. Parameters in a cancer screening MSM may include sensitivity and specificity of screening test by stage of disease at time of screening, timing of screening tests, compliance with screening, cancer incidence as a function of age and demographic variables, sojourn time distribution from pre‐clinical to clinical stages, and cause‐specific survival distributions by stage at diagnosis. The joint distribution of mortality and late‐stage incidence by arm (which we directly simulate in Section 4) may be derived from these MSM models. Second, published results for cancer screening trials report subject outcomes with different follow‐up times. Studies included in meta‐analyses estimate endpoints at varying follow‐up times. Late‐stage incidence reduction (Sl) and mortality reduction (Ml) parameters depend on follow‐up duration. If follow‐up time is limited, the full impact of screening on endpoints will not yet have occurred and Ml and Sl may be small in magnitude (close to 0). Ml is especially sensitive as cancer deaths take time to accrue following diagnosis. With very long follow‐up times, wash‐out can occur resulting in lower magnitude Ml and Sl as well. For example, if a screening program is performed annually for four years and endpoints are assessed many years (e.g., 20) later, most cancer deaths and late‐stage diagnoses result from tumors which began growing after the screening period concluded, where screening could not possibly be effective. Hu and Zelen [41] derived optimal follow‐up times for powering statistical tests to detect a mortality reduction due to screening. Sojourn distributions (time interval where disease is potentially detectable via screening test), screening test sensitivity, and survival distribution by stage all affected the relationship between mortality reduction and follow up time. Zhu et al. [42] have studied the effect of follow‐up time in the context of power calculations for screening trials using mortality and late‐stage incidence endpoints. These differences in follow‐up duration can alter the estimates of both primary and surrogate endpoints within each trial, potentially affecting the strength and consistency of correlations between primary and surrogate endpoints across trials. Meta‐analytic modeling has been used to study the quality of surrogates in many contexts, including therapeutics for cancer. Prasad et al. [25] reviewed 36 meta‐analyses of surrogates across many therapeutics and disease sites. Each of these 36 meta‐analyses performed meta‐analytic regression and reported the strength of correlation between a primary (e.g., overall survival) and surrogate (e.g., progression‐free survival) outcome. Many of these studies did not incorporate trial‐level estimate uncertainty, possibly influencing the strength of the estimated correlation between primary and surrogate endpoints. Adapting the simulations and theory developed here in Sections 4 and 5 to survival outcomes may lead to a better understanding of the effect of trial‐level uncertainty on meta‐analysis outside of the cancer screening setting considered in this work. For future studies, we recommend only using the meta‐analysis regression to validate surrogates when trial‐level uncertainty can be propagated into model fits, as done in methodology developed by [20] and [24]. When this is not possible, alternative assessments of surrogate quality may be useful, such as methods proposed in [17]. When performed, meta‐analytic modeling should include plots with size of trial‐level estimate uncertainty as we have done in Figure 3. Funding This work was supported by the National Cancer Institute (Grant Nos. P30CA016672, P50CA127001‐16, R01CA269696) and the National Center for Advancing Translational Sciences (Grant No. UM1TR004906). Conflicts of Interest The authors declare no conflicts of interest. Acknowledgments J.P.L. received support from the National Cancer Institute and the National Center for Advancing Translational Sciences of the NIH [P30CA016672 and CCTS UM1TR004906], Y.S. is partially supported by NCI BRG (P30CA016672) and R01CA269696. NLST data were accessed under project NLST‐2696 https://cdas.cancer.gov/approved‐projects/2696/. Reparameterization of Model In Section 3, the model is parameterized using conditional probabilities, e.g., pD|EC is the probability of death given early‐stage diagnosis in the control group. For the purpose of deriving asymptotic results, it is simpler to parameterize the model in terms of joint probabilities of occurrence of both primary mortality endpoint (death from cancer, no death from cancer) and surrogate endpoint (no cancer diagnosis, early‐stage diagnosis, late‐stage diagnosis). There are a total of six events (2 mortality outcomes × 3 diagnosis outcomes) per arm. However, since it is not possible for an individual to die of cancer without being diagnosed with cancer, only five of these events have non‐zero probability. Let pjkC be the probability of mortality endpoint j (j=0 is no death from cancer, j=1 is death from cancer) and diagnosis endpoint k (k=0 is no cancer diagnosis, k=1 is early‐stage diagnosis, and k=2 is late‐stage diagnosis) in the control arm. Identical definitions apply to the screen arm with superscript C replaced by S. One can map the conditional parameterization in Section 3 to the joint parameterization with Estimators are Proof of Theorem 1 Define p=(p11C,p02C,p12C,p11S,p02S,p12S)T and p^=(p^11C,p^02C,p^12C,p^11S,p^02S,p^12S)T. By the Central Limit Theorem where The parameters of interest (late‐stage reduction and mortality reduction) are By application of the Delta Method, where Proof of Theorem 2 We show Cov(S^,M^)=∑SM,12>0. From Equation (A2), we have Note that Using these derivations and the asymptotic covariance of p^ derived in Equation (A1), we have Note that r,t<0 and s,u>0. Thus rt>0 and su>0. Thus, to prove ∑SM,12>0, it is sufficient to show that A,B>0. Since A and B are defined analogously for control and screen, sufficient to show A>0. Note that pD|EC=p11C/(p11C+p01C) and pD|LC=p12C/(p12C+p02C). Thus, the theorem assumption is equivalent to which implies where the final inequality follows from Inequality (A3). Confidence Region for (S,M) A (1−α)100% Wald Confidence Region for (S,M) is where q2α is the (1−α) quantile of a chi‐squared distribution with 2 degrees of freedom and ∑^SM is a consistent estimator of ∑SM (defined in Equation (A2)). Note that ∑SM consists of three scalar quantities (Var(S^), Var(M^), and ρ). The quantity ρ cannot be estimated from data available in meta‐analysis. However Var(S^) and Var(M^) can be estimated. Specifically, recall that where pDS is the probability of mortality in the screen arm and pDC is the probability of mortality in the control arm. By the CLT By application of the delta method with M=f(pDC,pDS)=1−pDS/pDC, By application of the delta method with S=f(pSC,pSS)=1−pSS/pSC, Data Availability Statement The data that support the findings of this study are openly available in surrogate‐code at https://github.com/longjp/surrogate‐code. References Associated Data Data Availability Statement The data that support the findings of this study are openly available in surrogate‐code at https://github.com/longjp/surrogate‐code.