When caught early, before it spreads, cancer has a much better prognosis than when it is discovered in the advanced, metastatic stages of the disease. For instance, localized colorectal cancer has a ten-year survival of about 85%, but if the disease is metastatic, this figure drops to less than 10%. Thus, the goal of cancer screening (such as colonoscopy in the case of colorectal cancers) is to detect cancer as early as possible, before the onset of signs or symptoms. This topic is so important that we will be devoting an upcoming “Ask Me Anything” episode to all things cancer screening.
Yet despite cancer screenings’ benefit of averting certain cancer-related deaths, some are against population-level cancer screening in average-risk populations. We will weigh the collective arguments for and against cancer screening in the upcoming AMA, but we’d first like to zero in on some of the latest fuel to the anti-screening fire: a recently published study by Bretthauer et al. claiming that most recommended forms of cancer screening don’t extend lifespan. While this initially sparked alarmist headlines (“Cancer screening doesn’t extend lifespan!?”), the findings (or perhaps lack thereof) need to be put into a more nuanced context.
What did the study find?
Epidemiology studies often show that cause-specific mortality decreases as a population’s compliance with cancer screening increases, but observational studies are riddled with biases, such as healthy user bias (those who are generally more health-conscious are more likely to get screened) or socioeconomic bias (those with higher socioeconomic status have better access to healthcare and screenings). To avoid these confounds, the meta-analysis by Bretthauer et al. focused on randomized clinical trials to assess effects of screening on lifespan extension. Specifically, the study aimed to estimate the potential benefit of screening (i.e., catching cancer early) relative to the potential harms, as the majority of those who undergo screening will not have cancer, and cancer screening itself is not without risk. (Life-shortening complications, though rare, can occur with some screening procedures or from procedures that might follow a positive screening result.).
The investigators included up-to-date meta-analyses and individual RCTs for six common cancer screening modalities: mammography for breast cancer; fecal occult blood testing (FOBT), flexible sigmoidoscopy, and colonoscopy for colorectal cancer; prostate-specific antigen (PSA) testing for prostate cancer; computed tomography (CT) scanning for lung cancer in current or former smokers; and Papanicolaou (Pap) test cytology for cervical cancer. All included trials compared a screening group versus a no-screening control group over a follow-up period of 10-15 years.
From the 18 trials (over 2 million patients) that met the inclusion criteria, the meta-analysis reported that none of the cancer screening modalities resulted in a statistically significant reduction in all-cause mortality (ACM) relative to no-screening controls. Further, only flexible sigmoidoscopy was found to extend lifespan by a statistically significant margin, and even this “extension” was fairly short at 110 days (<4 months).
The importance of study design
Before starting any clinical study, the estimated effect of the intervention is used to calculate the number of study participants needed to show a statistically significant effect – i.e., the investigators must ensure that the study is sufficiently powered. Most cancer screening trials are powered to show a reduction in cancer-specific deaths. That is, a mammography study is powered to show a decrease in deaths specifically caused by breast cancer – not all-cause mortality. And yet, the meta-analysis by Bretthauer et al. primarily assessed changes in ACM – a metric cancer screening studies are not designed for.
To illustrate the impact of this issue, consider one of the most successful cancer screening studies to date, which was conducted in Sweden in the 1970s and 80s. This trial demonstrated that regular mammograms resulted in >30% reduction in breast cancer mortality relative to the no-screening group. Yet even with the significant reduction in breast cancer mortality, the investigators noted that a nine-year follow-up period wouldn’t have been long enough to show a change in ACM, since breast cancer accounted for only a small fraction (~7% at the time) of total deaths in women, meaning a 30% reduction in breast cancer mortality would at most change ACM by only 2%, well within the margin of error, and not an effect that could be confidently concluded.
More broadly, using ACM to determine the success of cancer screening is a nearly impossible task, at least over the period of time such studies are done. Another article published in the same issue estimated the number of participants required to reveal a 25% reduction in cancer-specific death, as well as the completely unrealistic population required to reliably show a reduction in ACM based on the mortality reduction from a single type of cancer, a number on the order of millions of participants. For example, if the true effect of breast cancer screening relative to no screening is a 25% reduction in breast cancer mortality, a trial would need to have 96,000 participants to show a significant reduction in breast cancer mortality, but the same trial would need more than 1.7 million participants to show a reduction in ACM, a number of subjects not achieved even by pooling multiple trials in the meta-analysis. For less deadly cancers, such as prostate cancer, the number of participants increases to more than 11 million study subjects. This means the expected outcome of any cancer screening study or even meta-analysis is highly unlikely to show reductions in ACM.
While Bretthauer and colleagues do report absolute differences in target cancer mortality between screening and no-screening, they do not provide statistical analyses on these figures to indicate whether effects are significant. However, with screening, target cancer mortality decreased for all screening modalities with the exception of the PSA test, a very low-specificity test for prostate cancer. Even in the absence of statistics, this trend is noteworthy, particularly when we consider that the average 10 to 15-year follow-up is likely still too short to evaluate cancer mortality in a population that is cancer-free at baseline. Longer trials, though difficult and expensive to conduct, would be necessary, especially for slow-developing cancers like colon cancer or for cancers that are often very responsive to treatment, like certain types of breast cancer. Over these short-duration trials, it would be difficult to show large differences in the effect on lifespan for target cancer mortality, let alone a difference in all-cause mortality (ACM).
Why this hasn’t changed my views of the utility of cancer screening
In addition to evaluating a nearly impossible outcome, the meta-analysis by Bretthauer and colleagues is filled with other problems that further limit, if not invalidate, the findings. The inclusion criteria were such that no more than five studies were pooled for analysis for any form of cancer screening; for colonoscopy, only one trial was included, which doesn’t add any new insight and just regurgitates what was found in that one study. The meta-analysis also included trials such as the Canadian Breast Cancer Screening trial, which has come under heavy scrutiny for failing to meet the standards of randomized trials and tampering with the process of randomization.
One of the other flaws of this study was its sole use of an intention-to-treat (ITT) analysis. In any randomized trial, it is rare for 100% of the intervention group to comply with the screening protocol, but an ITT includes the entire group regardless of participation. In addition to lapses in screening among the intervention group, non-compliance can also affect the control group, meaning that some in the non-screening groups may have gotten screenings. The purpose of using an ITT analysis is to avoid some of the biases common to epidemiology, but it also simultaneously (and artificially) lowers the reported efficacy of screening, an effect that increases as compliance decreases. For instance, in the only colonoscopy randomized trial (NORDICC), compliance was only 42%, meaning that an ITT analysis (a measure of effectiveness of an invitation to screen rather than efficacy of screening itself) would be highly unlikely to show a significant reduction in colon cancer death.
The efficacy of screening can be evaluated in the subset of compliant participants (known as a per-protocol analysis). However, non-compliance is known to be non-random based on biases of health consciousness or existing risk factors, so per-protocol analyses are known to have similar biases as observational studies. Therefore, a thorough investigation will include both a per-protocol and ITT analysis, and the magnitude of the difference between them helps to provide a sense of how strongly their respective flaws impacted results.
The bottom line
With the exception of some types of cancer (e.g., some prostate cancer) known to be less aggressive than others, individuals have a much better chance of surviving treatment and extending lifespan when cancer is detected early. Only a few forms of cancer even have widely available screening tests, so while the risks of possible harm should be weighed against the potential benefits for the individual, the best chance of not dying from cancer comes from catching cancer early through screening. Blanket statements that cancer screening doesn’t extend lifespan misrepresent methodological limitations and are harmful to public perception of the value of cancer screening in early cancer detection.
For a list of all previous weekly emails, click here.