When to Use Heart Failure Survival Analysis

You're analyzing a clinical trial with 299 heart failure patients. After 6 months of follow-up, 96 patients have died and 203 are still alive. Your instinct might be to run logistic regression: death (yes/no) predicted by ejection fraction, serum creatinine, and age. But you'd be throwing away critical information.

Here's what you're missing: Patient A died after 4 days. Patient B died after 285 days. Logistic regression treats both the same—just "death = yes." But the time dimension contains the signal you need. Patient A's rapid decline suggests acute decompensation. Patient B survived nearly the entire follow-up period. These are fundamentally different clinical trajectories, and survival analysis is the only framework that captures this.

Before we build models, let's establish what makes this a survival analysis problem rather than a standard classification task. The question isn't just "did they die?" but "when did they die, and what factors accelerated or delayed that outcome?"

The Three Features That Make This a Survival Problem

Survival analysis exists because clinical data has three characteristics that standard regression can't handle properly. Understanding these features determines whether you need survival methods or can use simpler approaches.

Feature 1: Time-to-Event is the Outcome

You're not predicting a binary outcome (dead vs alive) or a continuous measurement (ejection fraction percentage). You're predicting time until an event occurs. Each patient contributes a duration: 4 days, 115 days, 285 days. This time variable is the dependent variable in your model.

The heart failure dataset tracks "time" (follow-up days) and "DEATH_EVENT" (1 = died, 0 = survived). Together, these create your survival object. A patient with time=50 and DEATH_EVENT=1 experienced the event at 50 days. A patient with time=200 and DEATH_EVENT=0 was still alive at their last follow-up (200 days) but we don't know what happened after.

Feature 2: Censoring is Unavoidable

Here's the problem that makes survival analysis necessary: not every patient experiences the event during your observation window. Of the 299 heart failure patients, 203 were still alive when the study ended. We know they survived at least until their last visit, but we don't know their ultimate survival time.

This is called right-censoring, and it's not missing data. It's partial information. A patient followed for 180 days who didn't die tells you: "This patient survived at least 180 days." That's valuable information that survival analysis incorporates through censoring indicators.

Standard methods fail here. If you exclude censored patients, you lose 68% of your sample (203 out of 299) and introduce severe bias—you're only analyzing the sickest patients who died quickly. If you treat censored patients as "alive forever," you underestimate mortality risk. Survival analysis handles censoring correctly by using partial likelihood methods.

Feature 3: Hazard Varies Over Time

The risk of death isn't constant. In heart failure patients, mortality risk is typically highest in the first 30-60 days after diagnosis or hospital discharge, then decreases as survivors stabilize. This time-varying risk is called the hazard function.

Survival analysis models the hazard—the instantaneous risk of death at time t, given survival up to that point. This captures the reality that a patient who survives 6 months has already demonstrated some resilience; their remaining risk differs from a newly diagnosed patient. Logistic regression assumes risk is constant over the observation period, which is rarely true in clinical settings.

What Makes Heart Failure Data Ideal for Survival Analysis

Not every clinical dataset benefits from survival methods. Before we discuss how to build the analysis, let's identify what makes the heart failure dataset particularly well-suited for this approach. These criteria help you decide if your own data warrants survival analysis.

Clear Event Definition and Timing

The dataset tracks a discrete, unambiguous event: death. There's no gray area—you can't be "sort of dead." And the timing is precise, recorded in days from baseline to either death or last follow-up. This clarity is essential. Survival analysis struggles with vague endpoints like "disease progression" (what counts?) or imprecise timing (event occurred "sometime in Q2 2025").

For your own data, ask: Can I define the event in a way that two clinicians would agree on every case? Do I know the event time within a reasonable precision (days, weeks)? If not, you'll need to refine your endpoint definition before running survival analysis.

Sufficient Event Rate

With 96 deaths out of 299 patients, the event rate is 32%. This is high enough to produce stable estimates but not so high that everyone experiences the event (which would make censoring irrelevant). As a general rule, you want at least 20-30 events total for basic Kaplan-Meier analysis, and 10-15 events per predictor for Cox regression.

What's your sample size requirement? It's not about total patients—it's about events. If you have 1,000 patients but only 15 died, you can't fit a Cox model with 8 predictors. You'll get unstable hazard ratios and wide confidence intervals. The heart failure dataset's 96 events supports a Cox model with roughly 6-9 variables, which aligns with the 13 available clinical measurements.

Meaningful Follow-Up Duration

Follow-up ranges from 4 to 285 days, with a median around 130 days. This range is long enough to observe survival differences between risk groups. If everyone were followed for just 7 days, there wouldn't be enough time for clinical factors to differentiate survival curves. If follow-up were 10 years, you'd need to account for competing risks (patients dying from other causes).

For heart failure specifically, 6-9 months is clinically meaningful—it captures the critical post-diagnosis or post-hospitalization period when interventions matter most. Your follow-up window should match your research question. If you're studying 5-year cancer survival, 6-month follow-up won't answer the question.

Clinical Predictors Measured at Baseline

The dataset includes 13 baseline measurements: age, sex, ejection fraction, serum creatinine, serum sodium, and other lab values. These are measured before the follow-up period starts. This temporal ordering is crucial for causal interpretation—predictors must occur before the outcome.

Time-varying covariates (measurements that change during follow-up) require more complex models. The heart failure dataset keeps it clean: all predictors are fixed at baseline, so standard Cox regression applies. If your data includes measurements taken during follow-up (e.g., creatinine measured monthly), you'll need time-dependent covariate models.

How the Analysis Actually Works: From Data to Hazard Ratios

Let's walk through the statistical machinery that turns 299 patient records into actionable clinical insights. I'll focus on what the methods are doing and how to verify they're working correctly.

Step 1: Kaplan-Meier Estimation of Survival Functions

The Kaplan-Meier estimator answers: "What percentage of patients survive to time t?" It builds the survival curve by calculating survival probability at each time point where an event occurs.

Here's the logic: At time 0, 100% of patients are alive (S(0) = 1.0). The first death occurs at day 4. At that moment, 1 out of 299 patients has died, so the survival probability drops to 298/299 = 0.9966. The next death occurs at day 5. Now 2 of the original 299 have died, but we calculate the conditional probability: given you survived to day 5, what's the probability you survive past day 5? This is 297/298. The cumulative survival probability at day 5 is 0.9966 × (297/298) = 0.9933.

This process continues for every death event. The survival curve is a step function—it drops at each event time and stays flat between events. Censored observations don't create steps; they reduce the number of patients "at risk" for subsequent calculations. If 50 patients are censored at day 100, the denominator for deaths after day 100 decreases by 50.

The median survival time is the point where S(t) = 0.50. If the curve never drops below 50% during your follow-up period (because most patients survived), the median survival is undefined or reported as "not reached." This is common in cancer trials with effective treatments.

Step 2: Stratified Survival Curves to Identify Risk Factors

Next, we split the cohort by clinical variables and compare survival curves. For example, divide patients into low ejection fraction (<40%) versus normal ejection fraction (≥40%). Calculate separate Kaplan-Meier curves for each group. If the curves diverge, ejection fraction affects survival.

The log-rank test quantifies this divergence. The null hypothesis: both groups have identical survival distributions. The test statistic compares observed versus expected deaths in each group at each time point. If low-EF patients die more often than expected and normal-EF patients die less often than expected, you reject the null hypothesis (p < 0.05 means statistically significant difference).

In the heart failure data, you'd typically stratify by:

• Ejection fraction (reduced vs preserved)
• Serum creatinine (elevated vs normal, indicating kidney function)
• Age (above vs below median)
• Serum sodium (hyponatremia vs normal)
• Anemia (present vs absent)

Each stratification produces two or more curves. If all five clinical factors show significant log-rank tests (p < 0.05), you have univariate evidence that they affect survival. But these are unadjusted comparisons—they don't account for confounding. That's where Cox regression comes in.

Step 3: Cox Proportional Hazards Regression

Cox regression models the hazard function—the instantaneous risk of death at time t. The model is:

h(t) = h₀(t) × exp(β₁X₁ + β₂X₂ + ... + βₚXₚ)

Where h₀(t) is the baseline hazard (the hazard for someone with all predictors = 0) and the exponential term captures how covariates multiply that baseline hazard. The key assumption: the baseline hazard can be any shape (it's not parametrically specified), but the effect of covariates is to scale it proportionally.

The proportional hazards assumption means the hazard ratio between two patients stays constant over time. If Patient A has serum creatinine = 2.0 and Patient B has serum creatinine = 1.0, and the hazard ratio is 1.8, then Patient A's hazard is 1.8 times Patient B's hazard at every time point. The curves can have different shapes, but the multiplicative factor (1.8) doesn't change.

The coefficients (β values) are estimated using partial likelihood—a clever method that doesn't require knowing the baseline hazard h₀(t). For each event time, the method asks: "Among all patients still at risk, which one actually died?" It compares the covariate values of the patient who died to all patients who could have died at that moment. This process repeats for every death event, and the β values that maximize this partial likelihood are your parameter estimates.

Step 4: Interpreting Hazard Ratios

The exponentiated coefficients (exp(β)) are hazard ratios. If serum creatinine has β = 0.59, then HR = exp(0.59) = 1.80. Interpretation: for each 1-unit increase in serum creatinine, the hazard of death increases by 80%, holding all other variables constant.

A 95% confidence interval for the hazard ratio tells you the precision of this estimate. If HR = 1.80 with 95% CI [1.35, 2.41], you can be 95% confident the true hazard ratio is between 1.35 and 2.41. Because the interval excludes 1.0, the effect is statistically significant. If the CI were [0.90, 2.70], it includes 1.0, meaning the effect is not statistically significant—you can't rule out that serum creatinine has no effect.

For continuous variables, hazard ratios are per-unit increase. If age has HR = 1.05 per year, a 10-year age difference corresponds to HR = 1.05¹⁰ = 1.63 (63% higher hazard). For binary variables (anemia: yes/no), the HR compares the two groups directly. If anemia has HR = 1.42, patients with anemia have 42% higher hazard than those without.

Step 5: Checking the Proportional Hazards Assumption

Before you trust your Cox model results, verify the proportional hazards assumption. The standard diagnostic: Schoenfeld residuals. For each covariate, you get a residual at each event time. If the assumption holds, these residuals should have no trend over time—they should scatter randomly around zero.

Plot scaled Schoenfeld residuals versus time for each predictor. A flat line (zero slope) means the assumption holds. An upward or downward trend means the hazard ratio changes over time, violating the assumption. Run a formal test: correlate the residuals with time. A significant correlation (p < 0.05) indicates violation.

What do you do if the assumption fails? Three options:

1. Stratify on the violating variable. If sex violates proportional hazards, fit separate baseline hazards for males and females but keep shared coefficients for other predictors.
2. Include time interactions. Add a term like (serum_creatinine × time) to allow the effect to vary over the follow-up period.
3. Use restricted follow-up. If the assumption holds for 0-90 days but fails after day 90, analyze only the first 90 days where the model is valid.

In the heart failure dataset, ejection fraction and serum creatinine typically satisfy proportional hazards, but it's worth checking every variable. Don't skip this step—it determines whether your hazard ratios are interpretable.

Step 6: Model Performance with Time-Dependent ROC

How well does your Cox model predict outcomes? Standard ROC curves don't work for survival data because the outcome isn't binary—it's time-to-event with censoring. Instead, use time-dependent ROC curves that evaluate discrimination at specific time points.

At day 90, classify patients as "high risk" or "low risk" based on their predicted hazard from the Cox model. Then check: among patients who died by day 90, what percentage were correctly classified as high risk (sensitivity)? Among patients who survived past day 90, what percentage were correctly classified as low risk (specificity)? Plot sensitivity vs (1 - specificity) to get the ROC curve at day 90.

The area under the curve (AUC) quantifies discrimination. AUC = 0.5 means the model is no better than random guessing. AUC = 0.7 means moderate discrimination. AUC ≥ 0.8 means strong discrimination. For clinical models, AUC = 0.75-0.85 is typical for well-validated risk scores.

Generate time-dependent ROC curves at multiple time points: day 30, day 90, day 180. If AUC decreases over time, the model predicts short-term outcomes better than long-term. If AUC increases, early deaths are harder to predict (perhaps they're driven by unmeasured acute factors), but the model captures long-term risk well.

The Report Outputs You'll Actually Use

Now let's talk about what a complete survival analysis report looks like in practice. These are the specific tables and figures you need to make clinical decisions or publish results.

Overall Survival Curve

The primary output: a Kaplan-Meier curve showing survival probability from baseline to maximum follow-up (285 days in the heart failure data). The y-axis is survival probability (0 to 1), x-axis is time in days. The curve starts at 1.0 and steps down at each death event.

Key numbers to extract:

• Median survival: The time at which 50% of patients have died (if reached during follow-up)
• Survival at specific landmarks: S(30 days), S(90 days), S(180 days)
• 95% confidence intervals: The shaded region around the curve showing estimation uncertainty

For the heart failure dataset, you might find: 1-month survival = 0.89 [95% CI: 0.85-0.93], 6-month survival = 0.65 [95% CI: 0.59-0.71]. This tells you 89% of patients survive the first 30 days, but by 6 months, only 65% remain alive.

Stratified Survival Curves by Key Risk Factors

Multiple curves on one plot, stratified by clinically important variables. Typically you'll produce 4-6 stratified plots:

• Ejection fraction: Low (<40%) vs normal (≥40%)
• Serum creatinine: Elevated (>1.5 mg/dL) vs normal (≤1.5 mg/dL)
• Age group: <65 years vs ≥65 years
• Serum sodium: Hyponatremia (<135 mEq/L) vs normal (≥135 mEq/L)
• Anemia: Present vs absent

Each plot includes a log-rank test p-value. If p < 0.05, the curves are significantly different—the stratifying variable affects survival. If curves cross during follow-up (one group does worse early, better later), this suggests a violation of proportional hazards for that variable.

Cox Regression Results Table

A table with one row per predictor and columns for:

• Coefficient (β): The log-hazard ratio
• Hazard Ratio (HR): exp(β), the multiplicative effect on hazard
• 95% CI: Confidence interval for the hazard ratio
• p-value: Statistical significance (p < 0.05 = significant)

Example row:

Variable              β      HR    95% CI       p-value
Serum creatinine    0.59   1.80  [1.35-2.41]  <0.001
Age (per year)      0.048  1.05  [1.03-1.07]  <0.001
Ejection fraction  -0.034  0.97  [0.95-0.99]   0.003
Anemia (yes vs no)  0.35   1.42  [0.96-2.11]   0.08

Interpretation: Serum creatinine is strongly associated with mortality (HR=1.80, p<0.001). Each additional mg/dL increases hazard by 80%. Age also predicts mortality—each year adds 5% to hazard. Ejection fraction is protective—higher values reduce hazard by 3% per percentage point. Anemia shows a trend but isn't statistically significant (p=0.08, CI includes 1.0).

Proportional Hazards Diagnostic Plots

One plot per covariate, showing scaled Schoenfeld residuals versus time. Look for flat trend lines (slope ≈ 0). A systematic upward or downward slope indicates violation. The plot includes a smoothed line (loess or spline) and a p-value from the correlation test.

If all p-values > 0.05, the proportional hazards assumption holds for your model. If one or two variables violate it, consider the remedies discussed earlier (stratification, time interactions, or restricted follow-up).

Time-Dependent ROC Curves

Plots showing model discrimination at day 30, day 90, and day 180. Each curve plots sensitivity vs (1-specificity), with AUC values annotated. For example:

• Day 30 AUC = 0.82 [0.75-0.89]
• Day 90 AUC = 0.79 [0.73-0.85]
• Day 180 AUC = 0.76 [0.70-0.82]

The model performs best at 30 days (strong discrimination, AUC=0.82) and slightly worse at 180 days. This pattern suggests the clinical predictors capture acute risk factors more reliably than long-term trajectory.

Four Mistakes That Invalidate Your Results

Most survival analysis errors come from misunderstanding censoring, violating assumptions, or misinterpreting hazard ratios. Here are the critical pitfalls and how to avoid them.

Mistake 1: Treating Censored Patients as "Failures" or Excluding Them

I've seen analysts code censored patients (DEATH_EVENT=0) as "alive forever" and run standard logistic regression. This underestimates mortality because you're treating a patient who was last seen alive at 50 days the same as one who survived the full 285-day follow-up. Conversely, excluding censored patients throws away 68% of the sample and creates severe selection bias—you're only analyzing patients who died quickly.

The correct approach: use the survival object (time, event indicator) in every analysis. Kaplan-Meier and Cox models handle censoring through the risk set calculation—censored patients contribute to the denominator (patients at risk) up until their censoring time, then they're removed from subsequent calculations. This uses their partial information without making assumptions about what happened after censoring.

Mistake 2: Ignoring Proportional Hazards Violations

If you fit a Cox model without checking Schoenfeld residuals, you might report hazard ratios that aren't constant over time. For example, if ejection fraction has HR=0.95 in the first 30 days but HR=1.02 after 90 days (crossing from protective to harmful), reporting a single overall HR is misleading.

Always run the proportional hazards test. If p < 0.05 for a covariate, don't ignore it. Either stratify on that variable (allowing separate baseline hazards for each stratum), add a time interaction term (variable × time), or restrict your analysis to the time window where the assumption holds. Document which approach you used and why.

Mistake 3: Overfitting with Too Many Predictors

The heart failure dataset has 13 variables and 96 events. If you throw all 13 into a Cox model, you're fitting 7.4 events per variable—below the recommended 10-15. The result: unstable estimates, wide confidence intervals, and poor generalization. Your model might fit this dataset well but fail on new patients.

Before building the model, calculate events per variable (EPV). With 96 deaths, you can include about 6-9 predictors maximum. Use clinical knowledge and univariate analyses to select which variables to include. Prioritize variables with strong biological plausibility (ejection fraction, serum creatinine, age) and significant log-rank tests. Consider stepwise selection (backward elimination or forward selection based on AIC) to automate variable selection, but validate the final model on a holdout set or through cross-validation.

Mistake 4: Misinterpreting Hazard Ratios as Risk Ratios

A hazard ratio of 2.0 does not mean "twice the risk" or "twice the probability of death." It means the hazard—the instantaneous failure rate—is doubled. For rare events and short follow-up, this approximates a risk ratio, but it diverges for common events or long follow-up.

If you need actual risk predictions ("What's the probability this patient dies within 90 days?"), use the Cox model to generate predicted survival curves for individual patients, then read off S(90 days). The mortality risk at 90 days is 1 - S(90). This converts hazards into probabilities correctly, accounting for the baseline hazard shape.

How to Prepare Your Data for Survival Analysis

Before you can run Kaplan-Meier or Cox regression, your data must be structured correctly. Here's what the analysis pipeline expects and how to get there.

Required Data Structure

You need at minimum:

1. Time variable: The duration from baseline to event or censoring (in days, weeks, months—be consistent)
2. Event indicator: Binary variable (1 = event occurred, 0 = censored)
3. Baseline covariates: Predictors measured at or before the start of follow-up

The heart failure dataset already has this structure:

age, anaemia, creatinine, ejection_fraction, ..., time, DEATH_EVENT
75,  0,       2.0,        20,               ..., 4,    1
55,  0,       1.1,        38,               ..., 6,    1
65,  1,       1.5,        20,               ..., 7,    1
50,  1,       1.3,        35,               ..., 285,  0

Each row is a patient. The "time" column is follow-up duration in days. "DEATH_EVENT" is the indicator (1=died, 0=alive at last follow-up). All other columns are baseline clinical measurements.

Common Data Transformations

Before analysis, check for these issues:

Date conversion: If you have baseline_date and event_date (or censoring_date), calculate time = event_date - baseline_date. Make sure both dates are in the same format (YYYY-MM-DD) and calculate difference in days.

Missing covariates: Survival models typically require complete cases (no missing values in predictors). Options: (1) Impute missing values using median (continuous variables) or mode (categorical variables), or (2) use multiple imputation if missingness is substantial (>10%). Never delete patients with missing covariates unless missingness is rare (<5%).

Outlier handling: Clinical measurements can have extreme values due to lab errors. Check for biologically implausible values (creatinine = 50 mg/dL, age = 200 years). Either correct errors if you can verify true values, or winsorize extreme values (cap at 99th percentile) to reduce influence.

Variable scaling: Cox regression coefficients depend on variable units. If age is in years (range 40-95), β ≈ 0.05 per year. If age is in decades (range 4-9.5), β ≈ 0.50 per decade—same effect, different number. For interpretability, keep variables in clinically meaningful units (creatinine in mg/dL, ejection fraction in %, age in years).

Checking Data Quality

Before running survival analysis, verify:

• No negative times: Time must be ≥ 0. If you find negative values, your date subtraction is backwards or you have data errors.
• Events occur before censoring: If DEATH_EVENT=1, the patient must have time > 0 (you can't die before baseline). If DEATH_EVENT=0, time is censoring time.
• Consistent time units: Don't mix days and months. If baseline data uses days but follow-up uses months, convert to a single unit.
• All covariates measured at baseline: If a predictor is measured during follow-up (e.g., creatinine at day 30), it's a time-varying covariate. Standard Cox models assume baseline covariates only. Time-varying covariates require extended Cox models.

What the Results Tell You About Clinical Risk

Let's interpret a realistic set of results from the heart failure dataset to illustrate what you learn from survival analysis.

Overall Survival Pattern

The Kaplan-Meier curve shows survival probability declining from 100% at baseline to approximately 65% at 280 days. The curve drops steeply in the first 30-60 days (early mortality phase), then flattens (stable survivors). Median survival is around 210 days—half of patients survive beyond this point.

Clinical interpretation: Heart failure patients face highest mortality risk immediately after diagnosis or hospitalization. Survivors past 60 days have demonstrated some resilience, and their subsequent mortality rate is lower. This suggests early intervention and close monitoring in the first 2 months are critical.

Stratified Survival: Key Risk Factors

When you stratify by ejection fraction (<40% vs ≥40%), the curves diverge clearly. Low-EF patients have 50% survival at 150 days; normal-EF patients reach 50% survival at 260 days. Log-rank p < 0.001 confirms this difference is statistically significant. Similarly, elevated serum creatinine (>1.5 mg/dL) shows markedly worse survival than normal creatinine (p < 0.001).

Clinical interpretation: Reduced ejection fraction and impaired kidney function (high creatinine) are major mortality predictors. These aren't surprising—both reflect organ dysfunction. But the magnitude matters: a 110-day difference in median survival is clinically substantial. Patients with both risk factors (low EF and high creatinine) are in a very high-risk group requiring aggressive management.

Cox Model: Independent Predictors

The multivariable Cox model identifies variables that independently predict mortality after adjusting for confounders:

• Serum creatinine: HR = 1.80 [1.35-2.41], p < 0.001. Each mg/dL increase raises hazard by 80%. This remains significant even after adjusting for age, ejection fraction, and other variables.
• Age: HR = 1.05 [1.03-1.07], p < 0.001. Each additional year increases hazard by 5%. A 20-year age difference means HR = 1.05²⁰ ≈ 2.65 (2.65 times higher hazard).
• Ejection fraction: HR = 0.97 [0.95-0.99], p = 0.003. Each 1% increase in EF reduces hazard by 3%. Going from EF=25% to EF=40% (15-point increase) yields HR = 0.97¹⁵ ≈ 0.64 (36% hazard reduction).
• Serum sodium: HR = 0.95 [0.91-0.99], p = 0.02. Each mEq/L increase in sodium reduces hazard by 5%. Hyponatremia (low sodium) is a known poor prognostic sign in heart failure.

Variables that were significant in univariate analysis but dropped out in multivariable models (non-significant after adjustment) might include anemia, sex, or diabetes. This suggests their univariate effects were confounded by age or kidney function.

Clinical Risk Score

You can use the Cox model to create a risk score for new patients. For a 70-year-old with EF=30%, creatinine=2.0, and sodium=132, plug into the model:

Linear predictor = (0.048 × 70) + (-0.034 × 30) + (0.59 × 2.0) + (-0.05 × 132) = 3.36 - 1.02 + 1.18 - 6.60 = -3.08

Relative hazard = exp(-3.08) = 0.046 compared to a baseline patient (all covariates = 0, which is not realistic but serves as reference). More usefully, compare this patient to the average patient in your cohort or predict their survival probability at 90 days using the baseline hazard estimated from the Kaplan-Meier curve.

This risk stratification guides clinical decisions: high-risk patients (high creatinine, low EF, older age) might warrant more aggressive treatment, closer follow-up, or enrollment in disease management programs. Low-risk patients can be monitored less intensively.

Frequently Asked Questions

What makes survival analysis different from logistic regression for clinical outcomes?

Logistic regression tells you whether an event occurred (yes/no). Survival analysis tells you when it occurred. If you study 299 heart failure patients for 6 months and 100 die while 199 survive to the end, logistic regression uses only the 100 deaths. Survival analysis uses all 299 patients, treating the survivors as censored observations. This captures the full timeline of risk and handles incomplete follow-up correctly. The question isn't just "did they experience the event?" but "how long until the event occurred?"

When should I use Kaplan-Meier curves versus Cox regression?

Kaplan-Meier curves answer: "What's the survival probability over time for this group?" Use them to visualize survival patterns and compare 2-4 distinct groups (high vs low ejection fraction, treated vs untreated). Cox regression answers: "How do multiple risk factors simultaneously affect hazard?" Use it when you have continuous predictors (age, serum creatinine), multiple categorical factors, or need adjusted hazard ratios. In practice, start with Kaplan-Meier to understand univariate patterns, then build a Cox model to identify independent predictors.

How many events do I need for reliable Cox regression results?

The rule of thumb: 10-15 events per predictor variable. If you're fitting a Cox model with 5 predictors, you need 50-75 deaths (not 50-75 patients). With the heart failure dataset's 96 deaths, you can safely include 6-9 variables. Fewer events mean wider confidence intervals and unreliable hazard ratios. If you have 30 deaths and try to fit 10 predictors, your model will overfit and the results won't generalize. Calculate events-per-variable before building your model.

What does it mean when the proportional hazards assumption fails?

Cox regression assumes the hazard ratio between groups stays constant over time. If serum creatinine has HR=1.5 at 30 days, it should still be 1.5 at 180 days. When this fails, the effect changes over time. For example, ejection fraction might matter most in the first 60 days but become less important later. Check this assumption with Schoenfeld residuals. If violated, consider stratified Cox models (stratify on the violating variable) or time-varying coefficient models that allow the effect to change over follow-up periods.

Can I use survival analysis if not everyone was followed for the same duration?

Yes. That's exactly what survival analysis was designed to handle. In the heart failure dataset, some patients were followed for 285 days, others for just 4 days. This is called "variable follow-up" and survival methods handle it through censoring. A patient followed for 50 days who didn't die contributes information up to day 50, then is censored. Logistic regression can't do this—it would either exclude patients with short follow-up (losing data) or treat "alive at 50 days" the same as "alive at 285 days" (wrong). Survival analysis uses every patient's full timeline.

Related Articles