Why the Shape of Survival Curves Matters
A Kaplan–Meier plot can show whether groups’ observed survival experiences separate, approach one another, or reverse order over follow-up. These patterns matter in a group sequential study: at an interim look, they may prompt questions about whether a single treatment effect adequately describes the data. But a visual pattern is a clue, not a test. Sampling variation, censoring, and a small number of people still at risk can make a curve appear to change direction even when the underlying treatment effect does not.
The target concept is proportional hazards. Let \(\lambda_1(t)\) and \(\lambda_0(t)\) denote the event hazards at time \(t\) in the treatment and control groups. The hazards are proportional if their ratio is constant over time wherever the hazards are defined. Non-proportional hazards means that this ratio changes with time. A survival curve describes the probability of remaining event-free through a time; a hazard describes the instantaneous event rate among those still event-free. The two are related, but they are not interchangeable.
What the Curves Can and Cannot Show
Kaplan–Meier estimates are step functions. At each event time, the estimate is multiplied by the fraction of people at risk who do not have an event at that time. Censoring removes a person from later risk sets but does not itself produce a downward step. Under independent, non-informative censoring within each group, the estimate describes that group’s event-free survival over time.
Here \(t_j\) is an event time, \(d_j\) is the number of events at that time, and \(r_j\) is the number at risk immediately beforehand. A treatment comparison also needs an appropriate study design and analysis. Randomization supports a causal treatment comparison at the population level, but it does not make a noisy plotted difference conclusive.
Under proportional hazards, the survival functions satisfy \(S_1(t)=S_0(t)^\theta\). Since raising a number between zero and one to a fixed positive power preserves its ordering relative to the original number, the two population survival curves cannot repeatedly reverse order under this model. Thus, a clear crossing can be a warning sign. The reverse implication does not hold: curves that never cross do not establish proportional hazards. Hazards may vary over time while survival estimates remain in the same order.
Here \(H(t)\) is the cumulative hazard, with \(S(t)=\exp[-H(t)]\). The relation explains the ordering implication, not a visual test: estimated curves fluctuate, and departures from the model can be hard to see. In particular, divergence or convergence describes changing separation between survival estimates. It does not directly measure the hazard ratio. A growing gap may arise under proportional hazards, and a shrinking gap alone does not prove that hazards are non-proportional.
Reading Patterns Carefully
| Visual pattern | Reasonable interpretation | What it does not establish |
|---|---|---|
| Curves cross | Possible reversal in the observed survival ordering; investigate whether effects change over time. | That the population hazards definitely cross, or that a particular cause produced the pattern. |
| Curves diverge | The observed survival difference is growing over the displayed interval. | A changing hazard ratio; a constant hazard ratio can also yield increasing absolute survival differences. |
| Curves converge | The observed survival difference is shrinking over the displayed interval. | That the treatment effect has disappeared or that hazards are equal late in follow-up. |
| Curves overlap | The plotted estimates are similar at the displayed resolution. | Equivalence, absence of a treatment effect, or proportional hazards. |
Check where the pattern occurs and how much information supports it. A late crossing when only a few participants remain at risk is less persuasive than a sustained reversal during well-supported follow-up. Read the numbers at risk, censoring marks, time scale, and uncertainty alongside the lines. Wide uncertainty at late times means that the plotted tail should not be treated as a precise description of the population.
The Greenwood estimate gives a standard error for the Kaplan–Meier estimate when the relevant risk sets are nonempty and the estimate is positive:
It reflects uncertainty in a group’s estimated survival, not a formal test of whether the hazards are proportional. Separate confidence intervals for two curves also do not, by themselves, provide a calibrated test of a difference between curves or a time-varying hazard ratio.
Worked Example: A Visible Crossing Is a Warning, Not a Verdict
Suppose a hypothetical randomized study’s Kaplan–Meier output gives these survival estimates at three scheduled times. The values are illustrative, not results from a real study.
| Time (months) | Control \(\widehat S(t)\) | Treatment \(\widehat S(t)\) |
|---|---|---|
| 6 | 0.80 | 0.72 |
| 12 | 0.60 | 0.62 |
| 18 | 0.45 | 0.50 |
At 6 months, the treatment estimate is \(0.72-0.80=-0.08\), an 8-percentage-point lower estimated event-free survival. At 12 months, it is \(0.62-0.60=0.02\), a 2-point higher estimate. The ordering has reversed between those times, so the plotted curves must cross somewhere between them if the estimates are connected as step functions.
One descriptive check is to transform the estimates to cumulative hazards, \(\widehat H(t)=-\ln\{\widehat S(t)\}\). At 6 months, the treatment-to-control ratio of these cumulative-hazard estimates is \(-\ln(0.72)/[-\ln(0.80)]=0.3285/0.2231=1.472\). At 12 months it is \(-\ln(0.62)/[-\ln(0.60)]=0.4780/0.5108=0.936\). The changing ratios are consistent with a possible departure from a constant hazard ratio, but ratios of estimated cumulative hazards are not hazard ratios. Nor do two noisy estimates prove a population departure.
The appropriate conclusion is that the crossing warrants examination of uncertainty, numbers at risk, censoring, and a suitable prespecified formal analysis. It does not justify claiming from the plot alone that treatment is harmful early and beneficial later.
Worked Example: Divergence Does Not Automatically Mean Non-Proportional Hazards
In another hypothetical study, estimated control survival is \(0.90\) at 6 months and \(0.70\) at 12 months; treatment estimates are \(0.85\) and \(0.55\), respectively. The treatment-minus-control differences are \(0.85-0.90=-0.05\) at 6 months and \(0.55-0.70=-0.15\) at 12 months. The absolute gap has widened from 5 to 15 percentage points.
This is divergence in the plotted survival estimates, but it does not alone show a changing hazard ratio. Under a constant hazard ratio \(\theta=1.30\), for example, if control survival at a particular time is \(0.90\), the proportional-hazards relation gives treatment survival \(0.90^{1.30}=\exp\{1.30\ln(0.90)\}\approx0.872\). If control survival later is \(0.70\), it gives treatment survival \(0.70^{1.30}\approx0.629\). The absolute gaps are about 2.8 and 7.1 percentage points: they grow even though the assumed hazard ratio stays fixed at 1.30.
The hypothetical observed estimates are not identical to this illustration, and two time points cannot settle the model question. The calculation demonstrates the key distinction: a changing absolute separation between survival curves is not the same as a changing relative hazard. Review the full trajectories and uncertainty, then use a formal analysis appropriate to the planned estimand and design.
Worked Example: A Late Step Can Be Driven by a Tiny Risk Set
Consider a hypothetical treatment group with 4 people at risk just before a late event time and 1 event at that time. The Kaplan–Meier multiplier is \(1-1/4=0.75\). If the estimate just before this event is \(0.60\), then just after it the estimate is \(0.60(0.75)=0.45\), a visible drop of 0.15.
The Greenwood contribution from this event time is \(1/[4(4-1)]=1/12=0.0833\). If this is the only contribution up to the time shown, the estimated standard error just after the event is \(0.45\sqrt{0.0833}=0.130\). That large uncertainty relative to the estimate cautions against interpreting the step as compelling evidence of a late treatment-effect change.
If the other group has a higher estimate at this time, the lines might appear to cross. But with only four people at risk in the treatment group, the tail is fragile. Report the risk-set size and uncertainty, and avoid a confident claim based on this step alone.
From Visual Clue to Evidence
A sound review proceeds in stages. First, describe the observed pattern without attributing a cause: for example, “the treatment curve is lower early and higher later.” Next, check whether that pattern is supported by adequate numbers at risk and whether it persists over meaningful portions of follow-up. Then distinguish the descriptive observation from the inferential question: does a formal analysis provide evidence against the proportional-hazards assumption, with uncertainty and the study’s analysis plan taken into account?
Read the time axis, numbers at risk, censoring marks, and uncertainty; identify whether the apparent pattern depends on a sparse tail.
Record whether curves cross, diverge, converge, or overlap, and when. Treat this as a descriptive finding, not a conclusion about the true hazard ratio.
Assess proportionality with an appropriate prespecified diagnostic or model-based analysis, reporting its uncertainty and limitations.
At a planned interim look, curve inspection can help identify patterns worth discussing, but it must not become an unplanned route to an efficacy or harm claim. Group sequential boundaries calibrate repeated formal decisions for the prespecified monitoring statistic and analysis. They do not automatically calibrate decisions based on repeatedly inspecting curves or choosing a test after seeing their shapes. Preserve the monitoring plan, record exploratory observations as such, and have any consequential analysis reviewed under the protocol’s decision rules.
A formal diagnostic can itself be inconclusive. Limited events, short follow-up, heavy censoring, or a subtle time variation may leave little power to identify non-proportionality. Conversely, with enough information, a small departure may be statistically detectable but not important to the clinical interpretation. The curve is therefore useful for understanding what an average summary may conceal, but it is not a substitute for defining the estimand, checking assumptions, and evaluating uncertainty.
Common Mistakes and Limitations
- Equating a crossing with proof. A crossing of empirical estimates is a visual clue. Random variation can create one even if population hazards are proportional.
- Equating divergence with a changing hazard ratio. Absolute survival differences can grow when the hazard ratio is constant, as the worked calculation shows.
- Reading the tail as equally reliable. A late step based on a few people at risk can shift a curve sharply and carry substantial uncertainty.
- Calling overlapping curves equivalent. Similar plotted estimates do not prove that treatment effects are absent or clinically equivalent.
- Treating censoring as an event. Censoring does not make the Kaplan–Meier curve step down, but informative censoring can undermine the usual interpretation of its estimates.
- Using unplanned interim impressions as decisions. Repeated visual review does not inherit the error-rate control of prespecified group sequential boundaries.
The main limitation is that a survival plot is a noisy, marginal description, while the proportional-hazards condition concerns the underlying instantaneous hazards. Visual assessment helps flag possible departures and communicate their timing; it cannot establish the model, identify the cause of a pattern, or guarantee that one summary is adequate.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If the treatment survival curve crosses the control curve late, what does the crossing suggest, and what additional information should be checked?
- Can survival curves diverge under a constant hazard ratio? Explain briefly.
- A treatment curve drops at a time when 3 people are at risk and 1 event occurs. What is the Kaplan–Meier multiplier, and why should the step be interpreted cautiously?
- Does overlapping Kaplan–Meier curves prove equivalence between treatments? Why or why not?
Solutions. (1) The observed survival ordering reverses, which is a clue of possible time variation, not proof. Check risk-set sizes, censoring, uncertainty, and formal analysis. (2) Yes. The proportional-hazards relation can produce a growing absolute survival gap even with a fixed hazard ratio. (3) The multiplier is \(1-1/3=2/3\). With only three people at risk, the late estimate is unstable and has substantial uncertainty. (4) No. Similar estimates do not establish equivalence; that requires an appropriate design, estimand, and inferential procedure.