Q-Bank Breakdown: Kaplan-Meier curves — Why Every Answer Choice Matters

Kaplan–Meier (KM) curves look deceptively simple: two lines, a few drops, maybe some tick marks. But in a q-bank vignette, every visual detail is an answer choice waiting to happen—especially censoring, median survival, and the (often-misunderstood) log-rank test. Let’s walk through a classic clinical scenario and then dismantle each distractor the way the exam writers want you to.

Tag: Biostatistics > Study Design & Probability

---

The Clinical Vignette (Q-Bank Style)

A randomized controlled trial compares a new oral anticoagulant (Drug A) vs warfarin (Drug B) for prevention of recurrent venous thromboembolism (VTE) after an initial DVT. Patients are followed for 18 months. The primary outcome is time to recurrent symptomatic VTE.

A Kaplan–Meier plot is shown:

• The Drug A survival curve stays above Drug B throughout follow-up.
• Curves separate early and continue to diverge.
• Several tick marks appear on both curves.
• At 12 months, the Drug A curve is at 0.90 and the Drug B curve is at 0.80.
• The median survival is not reached for Drug A by 18 months; Drug B crosses 0.50 at ~16 months.

Question: Which statement best interprets the results?

Answer Choices

A. Drug A reduces recurrent VTE incidence at 12 months by 10 percentage points compared with Drug B.
B. The tick marks represent patients who had recurrent VTE at those times.
C. Because median survival is not reached for Drug A, the study cannot compare treatments.
D. A log-rank test is the most appropriate test to compare these time-to-event curves.
E. A chi-square test on 12-month outcomes is superior because it ignores censoring and avoids bias.

---

Correct Answer: D. A log-rank test is the most appropriate test to compare these time-to-event curves.

Why this is correct (what Step wants you to say)

• Kaplan–Meier curves display time-to-event data with censoring.
• The log-rank test compares the entire survival experience over time (i.e., compares observed vs expected events across time points).
• It appropriately incorporates censored observations.

High-yield KM interpretation points

• Y-axis: estimated probability of being event-free (“surviving”) up to time $t$.
• Each downward step: an event occurred (e.g., recurrent VTE).
• Censoring: patient left study, lost to follow-up, or study ended before event → contributes information up to censor time.

---

The Kaplan–Meier Essentials You Must Know (USMLE-Friendly)

What a Kaplan–Meier curve is (in one sentence)

A KM curve is a nonparametric estimate of the survival function $S(t)$ that accounts for right-censoring.

How KM calculates survival (conceptual)

At each event time:

• Let $n_i$ = number at risk just before time $t_i$
• Let $d_i$ = number of events at $t_i$
• Then the conditional probability of surviving past $t_i$ is $\left(1 - \frac{d_i}{n_i}\right)$
• Overall survival multiplies across event times:

$$S(t) = \prod_{t_i \le t}\left(1 - \frac{d_i}{n_i}\right)$$

“Tick marks” = censoring

A tick mark means:

• The patient did not have the event at that time
• They simply stopped being observed afterward

Median survival

• Median survival time is when $S(t)=0.5$
• If the curve never drops to 0.5, median survival is “not reached.”
• That does not invalidate comparisons.

---

Systematically Destroying Each Distractor

A. “Drug A reduces recurrent VTE incidence at 12 months by 10 percentage points…”

Why it’s tempting: The graph shows 0.90 vs 0.80 at 12 months → looks like an easy absolute risk reduction (ARR).

Why it’s not the best answer (and what’s technically wrong):

• KM estimates are time-to-event probabilities with censoring, not simple raw “incidence” unless carefully phrased.
• Saying “incidence” at 12 months implies a fixed-time risk comparison (which can be done), but the question asks for the best interpretation of the curves overall.

How to rescue this concept for Step:
If asked specifically for absolute risk reduction at 12 months, then:

• $ARR \approx 0.90 - 0.80 = 0.10$ (10 percentage points event-free difference)
  But you’d still remember: this is a KM-estimated difference, not necessarily the crude proportion.

---

B. “Tick marks represent patients who had recurrent VTE at those times.”

Why it’s wrong:

• Tick marks represent censoring, not events.
• Events are shown by the vertical drops.

High-yield rule:

• Drops = events
• Ticks = censored

Classic censoring reasons:

• Lost to follow-up
• Withdrawal
• Study ends before event
• Competing reasons for leaving observation (sometimes)

---

C. “Because median survival is not reached for Drug A, the study cannot compare treatments.”

Why it’s wrong:

• You can compare curves without median survival.
• Many effective therapies lead to “median not reached,” especially with short follow-up or low event rates.

What you can report instead:

• Log-rank p-value (difference across the curve)
• Hazard ratio (from Cox proportional hazards model)
• Survival probability at specific times (e.g., 12-month event-free survival)

---

E. “A chi-square test on 12-month outcomes is superior because it ignores censoring and avoids bias.”

Why it’s wrong (and kind of backwards):

• Ignoring censoring creates bias and throws away time information.
• A chi-square test at a fixed time point:
  • treats outcomes as simply “event by 12 months: yes/no”
  • may misclassify or exclude censored patients
  • loses information about when events happen

When a fixed-time comparison might be used (rarely on Step):

• If everyone has complete follow-up to 12 months (no censoring)
• Or if the endpoint is specifically “event by 12 months,” with complete ascertainment

Otherwise, for time-to-event with censoring → KM + log-rank, and often Cox regression.

---

Log-Rank vs Cox: Don’t Mix Them Up

Exam clue:

• If question asks “best statistical test to compare KM curves” → log-rank
• If it asks for “effect size while adjusting for confounders” → Cox

---

High-Yield Traps and Visual Clues on KM Curves

1) Early separation vs late separation

• Early separation suggests early benefit (or harm)
• Late separation could suggest delayed effect or insufficient early power

2) Crossing curves

If curves cross, proportional hazards may be violated → log-rank and Cox can be misleading.

3) Many tick marks late in follow-up

If lots of censoring late:

• the tail of the curve is based on fewer patients
• estimates become less precise

4) “Number at risk” table (if shown)

Always check it. If only a handful remain at risk late, don’t overinterpret late differences.

---

How USMLE Commonly Tests Kaplan–Meier

Know these “one-liners” cold:

• Kaplan–Meier: estimates survival over time with censoring
• Censoring: patient leaves study/event not observed; contributes info up to censor time
• Log-rank: compares KM curves (time-to-event)
• Median survival: time when $S(t)=0.5$; can be “not reached”
• Hazard ratio (Cox): relative instantaneous event rate; assumes proportional hazards

---

Quick Wrap-Up (What You’d Answer Under Time Pressure)

• The graph displays time-to-event with censoring → interpret with KM methods.
• Tick marks are censored, not events.
• The most appropriate comparison of the curves is the log-rank test.
• Fixed-time chi-square comparisons usually waste information and mishandle censoring.