Evaluating Screening for the Early Detection and Treatment of Cancer Without Using a Randomized Control Group
提出一种名为周期性筛查评估（PSE）的新方法，利用非随机对照数据估计筛查对癌症死亡率的影响，并通过乳腺癌数据验证其有效性。
Abstract Sometimes the only data available for evaluating screening for the early detection and treatment of cancer come from studies lacking a randomized control group. To evaluate these data, various approaches have been proposed. They include constructing a model to estimate the duration of preclinical cancer and the sensitivity of the screen, formulating a model to estimate rates of cancer mortality, and using a nonrandomized control group to estimate age-specific incidence and mortality rates. These methods have various limitations: Either the estimated quantities are insufficient for a complete evaluation, strong parametric assumptions are required, or an assumption must be made that the nonrandomized control group yields unbiased estimates of cancer incidence. To bypass some of these limitations, a new method has been developed for analyzing screening data without a randomized control group. The method is called periodic screening evaluation (PSE) because it requires data from periodic screens rather than a single screen. Using weakly structured parametric models, PSE estimates the effect of screening on cancer mortality, using only data from persons offered screening. PSE's major innovation is a technique for estimating age-specific cancer incidence using older screened persons as controls for younger screened persons. The limitations of PSE are often preferable to those associated with other methods. PSE's principal limitations are (a) use of a nonrandomized control group to estimate case-fatality rates in the absence of screening and (b) an assumption that, given a person's age, the year of birth adds no information for predicting cancer incidence. PSE was applied to data on persons offered screening for the early detection and treatment of breast cancer in which a randomized control group not offered screening was available for validation. Using only data from persons offered screening, the probability of averting death from cancer due to starting periodic screening at various ages was estimated. Confidence intervals were large. Nevertheless, a fivefold increase in sample size, which probably would have been feasible without the constraints of a randomized trial, would likely have yielded adequate precision. For validation purposes the cumulative number of observed deaths from cancer in the randomized control group was compared with the estimated cumulative number based on applying PSE only to women offered screening. The agreement was good. Key Words: EM algorithmObservational studySurvival analysis