Extending Trial Results to a Representative Target Population

Kosuke Inoue,William Hsu
DOI: https://doi.org/10.1001/jamanetworkopen.2023.46302
2024-02-01
JAMA Network Open
Abstract:Randomized clinical trials (RCTs) estimate the causal effect of an intervention on the outcome in a target population through careful design to control for potential biases. However, this effect varies based on how the target population is defined, and in practice, physicians are tasked with determining the relevance of RCT results for patient populations with different characteristics than the study population. Approaches that extend inferences from a study population to a representative target, such as transportability analysis, provide powerful tools to address questions such as who would most benefit from a clinical intervention. 1 ,2 Lung cancer remains the leading cause of cancer-related death worldwide, driven by the advanced disease at the time of diagnosis. While screening the general population using chest radiographs showed no benefit, evidence from 2 large trials showed lung cancer mortality benefit when screening high-risk populations based on age and smoking history using low-dose computed tomography (LDCT). The National Lung Screening Trial enrolled 53 454 participants randomly assigned to undergo 3 annual screens with either LDCT (n = 26 722) or chest radiography (n = 26 732). 3 The trial reported a 20% (95% CI, 7%-27%) relative risk reduction of lung cancer–related mortality in men and women compared with chest radiography at 6.5-year follow-up. The Dutch/Belgian randomized lung cancer screening trial (NELSON) reported its survival results involving 15 792 participants who were screened at 1, 3, and 6 years, showing 24% (95% CI, 7%-39%) relative risk reduction of lung cancer–related mortality in men undergoing LDCT screening as compared with no screening at 10-year follow-up. 4 These trials show compelling evidence for the studied populations that lung cancer screening using LDCT can reduce lung cancer mortality. However, compared with the general US population that would be eligible for screening, both NLST and NELSON participants were younger, had fewer comorbidities, were less likely to currently smoke, were more educated, and had more limited racial and ethnic diversity. Would a similar reduction in lung cancer mortality using LDCT vs chest radiographs be observed in the general population? The ability to extrapolate trial results to a target population is critical to enable the generalization of knowledge obtained from a trial and informing clinical decisions and public health policy. Robertson et al 5 estimated the effectiveness of LDCT vs radiography screening in a nationally representative target population to determine whether the estimated effect size differed from reported NLST findings. Their analysis leveraged survey data collected as part of the National Health Interview Survey (NHIS) in 2010, which was specifically selected given the availability of cancer-related health behavior, screening, and risk assessment data and the data having been collected before the US Preventive Services Task Force (USPSTF) issued a recommendation for lung cancer screening in 2013, avoiding the possibility that included patients may have already undergone lung cancer screening. A composite data set of NLST and NHIS participants was created, including data on screening strategy and outcomes (from the NLST only) and covariates commonly used in lung cancer risk models (from NLST and NHIS). Inverse odds weighting was used to reweight the results of the NLST to the target population by estimating the probability of trial participation and screening assignment in the trial. Based on this analysis, the relative rate reduction in lung cancer-specific mortality was 18% (95% CI, 1%-33%) vs 21% (95% CI, 9%-32%) reported in the NLST. To conduct their transportability analysis, the authors made 2 key assumptions. 5 First, to ensure conditional exchangeability over trial participation status, it is essential to comprehensively incorporate relevant demographic and clinical characteristics to construct the model for the likelihood of trial participation (ie, NLST vs NHIS). Second, every combination of covariates in the target population should possess a nonzero probability of being observed within the trial. Because transportability analysis does not allow us to extrapolate the results to a population that is not studied in the original RCT, we need to include a substantial number of participants from marginalized groups (eg, minoritized racial and ethnic groups, low-income populations) to facilitate the transportability of findings to the general population. In addition to these assumptions, when incorporating the results of transportability analysis into clinical guidelines or policymaking, the treatment or intervention's consistency (with the original trial's design) must be emphasized. Considering that the NLST took place in 2010, broadened criteria for LDCT screening (ie, adults aged 50 to 80 years -Abstract Truncated-
medicine, general & internal
What problem does this paper attempt to address?