How to Read a Cancer Survival Statistic

A cancer survival statistic answers a narrower question than most people hear in it: it reports the fraction of people still alive a fixed time after diagnosis, usually five years. It does not report how many were cured, how many were helped by treatment, or whether finding a cancer sooner changed the day anyone died. That distinction matters because a survival number can rise for reasons that have nothing to do with anyone living longer, and telling the two apart is the whole skill.

Key points#

• Survival measures time from diagnosis; screening changes when diagnosis happens, so it can inflate survival without saving a single life.
• Relative and net survival correct for deaths from other causes, but neither corrects for the moment the clock starts.
• The honest test of whether screening helps is a fall in cancer deaths per population (mortality), ideally in a randomized trial.
• A large relative benefit can sit on top of a small absolute one, so always ask for the underlying rates.

Survival and mortality are different measurements#

It helps to separate two numbers that sound alike. Survival follows a group of diagnosed patients forward and asks what share is alive later, while mortality looks at a whole population and asks how many die of the cancer per year. You can push survival up simply by changing who enters the diagnosed group or when they enter it. Mortality is far harder to fool, because it counts deaths across everyone whether or not they were ever diagnosed. Keep this pair in mind, because most of the confusion around screening comes from treating a survival gain as if it were a mortality drop.

Three versions of survival, from crude to refined#

Plain five-year survival is the simplest form. Diagnose 100 people, count 70 alive at year five, and survival is 70 percent. The trouble is that some of those deaths came from heart disease, other illness, or age, and had nothing to do with the tumor.

Two refinements try to remove that noise. Relative survival compares the diagnosed group against a matched slice of the general population of the same age, sex, and calendar year. If the cancer group reaches five years at 70 percent while a comparable cancer-free group would reach 90 percent, relative survival is the ratio, roughly 78 percent. Net survival goes one step further and models the survival that would be seen if the cancer were the only thing people could die of. A 2022 scoping review in the journal Cancers describes the Pohar Perme estimator, introduced in 2012, as the reference method for net survival, because it strips out background death rates in a way that lets populations with very different baseline mortality be compared on equal terms.

These are real gains. They let you set a young population beside an older one, or one country beside another, without ordinary differences in life expectancy pretending to be differences in cancer care. Yet every one of them still measures time starting at diagnosis, and that shared anchor is where the trouble begins.

Lead-time bias: an earlier start line, the same finish#

Picture a tumor that will end a life at age 70 regardless of what anyone does. Found because of symptoms at 67, the patient records three years of survival and counts as a death inside the five-year window. Now rewind and let a screening test catch the identical tumor in the identical person at 60. The patient still dies at 70. This time the record shows ten years of survival, and they clear the five-year mark as a success story.

The National Cancer Institute walks through this same scenario and states the point flatly: in its version, the patient does not live even a second longer. The disease did not change. Only the starting line moved. This is lead-time bias, and it is baked into any comparison built on survival. Detecting a cancer earlier lifts the survival statistic automatically, whether or not earlier detection did any good.

Length-time bias and overdiagnosis: screening favors the tame tumors#

The second distortion is quieter. Tumors grow at very different speeds. An aggressive cancer spends little time in a detectable but symptom-free phase, so a test run every year or two is unlikely to land during that brief window. A slow, indolent cancer sits in that phase for years, giving screening many chances to find it. The cancers a screening program collects are therefore weighted toward the least dangerous ones, which do well no matter how they are treated. That skew is length-time bias.

Its most extreme case is overdiagnosis, the discovery of a cancer that would never have caused symptoms or death within a person's natural lifespan. Every overdiagnosed case is guaranteed to be a five-year survivor, so each one nudges the statistic upward while helping nobody. The National Cancer Institute cites estimates that roughly 19 percent of screen-detected breast cancers, and somewhere between 20 and 50 percent of screen-detected prostate cancers, are overdiagnosed. Pile enough of these harmless findings into the denominator and, in the institute's own illustration, survival can march from 40 percent to 80 percent while the identical number of people die.

The measurement that resists the traps#

Stack the two biases together and the warning writes itself. A jump in five-year survival might signal genuine therapeutic progress, or it might reflect nothing more than an earlier clock and a denominator stuffed with cancers that were never going to hurt anyone. Survival by itself cannot separate these stories. Mortality can. Deaths from the cancer per unit of population, counted across the screened and the unscreened together, do not care when the clock started, and overdiagnosed cases never lower them. This is why the National Cancer Institute treats a reduction in cancer deaths within a randomized trial, rather than a rise in survival, as the trustworthy signal that screening works.

Reading the size of a benefit#

Even a real mortality benefit deserves a second look at its scale. The National Lung Screening Trial reported that low-dose CT screening cut lung-cancer deaths by 20.3 percent compared with chest radiography. That relative figure sounds enormous until you see the rates underneath it: about 247 lung-cancer deaths per 100,000 person-years in the CT arm against roughly 309 in the radiography arm. Both numbers describe the same finding. The relative reduction is the dramatic version; the absolute rates tell you how far the actual risk moved.

None of this makes survival statistics useless or screening a waste. A 2014 analysis in PLoS ONE by Maruvka, Tang, and Michor points to the opposite failure mode: waving away every survival gain as an artifact can bury real progress, because shifting incidence and detection can hide genuine treatment advances unless the analysis accounts for them. Once survival trends were normalized against incidence, that paper concluded, much of the rise in five-year survival reflected actual improvement in care rather than lead-time bias alone. The two lessons are halves of one idea. A survival number is where the questions start, not where they end. Before trusting one, ask which version it is, whether the compared groups were screened the same way, and whether anyone has shown a matching decline in deaths.