Cancer: Interpreting Changes in Relative Survival Over Time

by National Cancer Institute

(Page 2 of 2)

Increases in survival over time, however, even when based on data from tumor registries, such as SEER that include all cases in a given population, are difficult to interpret. They may reflect the benefits of early detection or improved treatment or both, but they may also result from lead-time bias and overdiagnosis, both of which occur commonly with screening.

Lead-time bias, which may result in longer survival of screen-identified cancers because the time before the cancer would have been diagnosed clinically, is included in the calculation of survival.

Overdiagnosis may result from finding cancers that would never have become manifest clinically and which might have a good prognosis. For example, autopsy series have shown a high percentage of occult early prostate carcinomas in elderly men who died of causes unrelated to prostate cancer. The discovery of these cancers through screening could increase the number of cases and give the appearance of stage shift, and of increases in survival or cure rates, without necessarily reducing mortality. An analysis of data reported by the SEER program for 1950 to 1996 found that changes over time in 5-year relative survival rates for 20 major cancers were essentially unrelated to trends in mortality rates for those cancers over the same period. The authors suggest that changes in 5-year survival rates are largely due to earlier diagnosis and to detection of subclinical cases that might never have surfaced clinically. They conclude that inferences about the effectiveness of early diagnosis or treatment should not be drawn from temporal changes in 5-year survival rates, but rather should be based on changes in mortality rates. Thus, changes in 5-year survival rates or stage shifts are not appropriate measures of the effectiveness of screening for early disease. Reductions in incidence rates for late-stage tumors represent a better measure of progress due to screening than 5-year survival trends, although such evidence is less compelling than reductions in mortality.

Study Designs

Varying study designs may be available to support a given summary. The strongest design would be obtained from a randomized controlled trial. It is, however, not always practical to conduct such a trial to address every question surrounding the field of screening. For each summary of evidence statement, the associated strength of study designs are listed. There are 5 study designs that are generally used in judging the evidence. In order of strength of design, the 5 levels are as follows:

1. Evidence obtained from randomized controlled trials.
2. Evidence obtained from nonrandomized controlled trials.
3. Evidence obtained from cohort or case-control studies.
4. Evidence obtained from ecologic and descriptive studies (e.g., international patterns studies, time series).
5. Opinions of respected authorities based on clinical experience, descriptive studies, or reports of expert committees.

Experimental trials are designed to correct for or eliminate selection, lead-time, length, healthy volunteer, and other biases when prospectively testing a detection procedure to determine its effect on health outcome. The highest level of evidence and greatest benefit from screening is mortality reduction in a randomized controlled trial. For most sites, such evidence is not available. Theoretically it is possible to conduct randomized trials for most interventions, but the sample size that is needed, the expense, and the duration of such trials for most cancers, frequently make this approach impractical. Therefore, evidence obtained by other methods is often used.

Case-control and cohort studies provide indirect evidence for the effectiveness of screening. Such evidence is particularly compelling for the effectiveness of screening for cervical cancer. Ecological correlation of mortality and intensity of screening has also been used in this context. Such studies do not prove a mortality-reduction effect, and the potential for bias to invalidate inferences from nonexperimental studies or to give misleading results, however, can be substantial.

Descriptive uncontrolled studies based on the experience of individual physicians, hospitals, and nonpopulation-based registries may yield some information about screening. The performance characteristics of various detection tests, such as sensitivity, specificity, and PPVs, are generally first reported in such descriptive studies. The first evidence that screening may be successful is an increase in the incidence of early cancers as well as a decreased incidence of late-stage metastatic cancers (stage shift); later, a reduction in deaths may occur. These descriptive studies do not establish efficacy because of the absence of an appropriate control group.

Disease-Specific and All-Cause Mortality Endpoints

Disease-specific mortality has been the most widely accepted endpoint in randomized clinical trials of cancer screening; however, the validity of this endpoint rests on the fundamental assumptions that the cause of death can be accurately determined and that the screening and subsequent treatment have negligible effects on other causes of death. Recent reviews of randomized clinical trials of cancer screening suggest that misclassification in cause of death has been a major problem and that misclassification has led to an overestimation of the effectiveness (or an underestimation of the harms) of screening. In contrast to disease-specific mortality, all-cause mortality depends only on an accurate ascertainment of deaths and when they occur and therefore is not affected by misclassification in cause of death. One major limitation of the all-cause mortality endpoint however is that it is unlikely to reveal a statistically significant effect of cancer screening because this intervention is usually targeted to a disease that causes only a small proportion of all deaths. Nevertheless, all-cause mortality should be considered in conjunction with disease-specific mortality to reduce the possibility that a major harm (or benefit) from screening is hidden by misclassification in cause of death.

Measures of Risk

Several measures of risk are used in cancer research. Absolute risk or absolute rate measures the actual cancer risk or rate in a population or subgroup (e.g., U.S. population, or whites or African Americans). For example, the SEER Program reports risk and rate of cancer in specific geographic areas of the United States.

Rates are often adjusted (e.g., age-adjusted rates) to allow a more accurate comparison of rates over time or among groups. The purpose of the adjustment is to make the groups more alike with respect to important characteristics that may affect the conclusions. For example, when the SEER Program compares cancer rates over time in the United States, the rates are adjusted to one age distribution. If this were not done, cancer rates would seem to increase over time simply because the U.S. population is getting older and the risk of cancer is higher in older age groups.

Relative risk (RR) compares the risk of developing cancer among those who have a particular characteristic or exposure with those who do not. RR is expressed as a ratio of risks or rates; it ranges from infinity to the inverse of infinity (i.e., zero). If the RR is greater than 1, the exposure or characteristic is associated with a higher cancer risk; if the RR is 1, the exposure and cancer are not associated with one another; if the RR is less than 1, the exposure is associated with a lower cancer risk (i.e., the exposure is protective). RR is often used in clinical trials of cancer prevention and screening to estimate the reduction in cancer risk or risk of death, respectively.

An odds ratio (OR) is often used as an estimate of the RR. It, too, indicates whether there is an association between an exposure or characteristic and cancer. It compares the odds of an exposure or characteristic among cancer cases with the odds among a comparison group without cancer. For relatively uncommon events/diseases such as a cancer diagnosis, it can be interpreted like a RR is interpreted; however, it becomes a progressively inaccurate estimate of the RR as the underlying absolute risk of an event/disease in the population under study rises above 10%. ORs are typically used in case-control studies to identify potential risk factors or protective factors for cancer.

Risk or rate difference (or excess risk) compares the actual cancer risk or rate among at least 2 groups of people, based on an important characteristic or exposure, by subtracting the risks or rates from one another (e.g., subtracting lung cancer rates among nonsmokers from that of cigarette smokers estimates the excess risk of lung cancer due to smoking). This can be used in public health to estimate the number of cancer cases that could be avoided if an exposure were reduced or eliminated in the population.

Population-attributable risk measures the proportion of cancers that can be attributed to a particular exposure or characteristic. It combines information about the RR of cancer associated with a particular exposure and the prevalence of that exposure in the population, and estimates the proportion of cancer cases in a population that could be avoided if an exposure were reduced or eliminated.

Number needed to screen estimates the number of people that must participate in a screening program for 1 death to be prevented over a defined time interval.

Average life-years saved estimates the number of years that an intervention saves, on average, for an individual who receives the intervention. This reflects mortality reduction as well as life extension (or avoidance of premature deaths).

About the Author

www.nci.nih.gov
The National Cancer Institute's research programs are extensive and contain many innovative initiatives. I invite you to explore our Web site to find out more about the exciting work being conducted here at NCI and by NCI-supported scientists throughout the country.