What Is Cox Proportional Hazards Regression?
Cox proportional hazards regression (usually just called "Cox regression") is a statistical method used to study how long it takes for an event to happen, and whether certain factors speed it up or slow it down. The "event" is often death, but it can be any clearly defined occurrence: disease recurrence, hospital readmission, graft failure, or recovery.
What makes Cox regression different from an ordinary regression model is that it uses both pieces of information every participant provides: whether the event happened, and exactly when (or, for participants who never experienced the event during the study, how long they were observed for without it happening). It answers a very specific, clinically useful question: "at any given moment, how does this factor change a patient's instantaneous risk of the event happening right now, compared to a patient without that factor?"
When Should Cox Regression Be Used?
Use Cox regression whenever your outcome has two components: an event indicator (did it happen — yes/no) and a time variable (how long until it happened, or how long the patient was followed if it didn't). It is the standard method for analyzing survival-type outcomes while simultaneously adjusting for multiple predictors — age, treatment group, tumor stage, comorbidities — in a single model.
| Clinical Scenario | Why Cox Regression Fits |
|---|---|
| Time to death after cancer diagnosis, by treatment arm | Event = death; time = months from diagnosis; adjusts for stage, age |
| Time to disease relapse after remission | Event = relapse; time = months in remission; adjusts for baseline risk factors |
| Time to hospital readmission after discharge | Event = readmission; time = days since discharge; adjusts for comorbidity burden |
| Time to graft failure after transplant | Event = graft failure; time = months post-transplant; adjusts for donor/recipient factors |
Cox Regression vs Logistic Regression
These two are the most commonly confused regression methods in medical research, because both can technically analyze a "yes/no" outcome. The key difference is what each one does with time.
| Feature | Cox Regression | Logistic Regression |
|---|---|---|
| Outcome | Time to event + whether it happened | Whether an event happened (yes/no) only |
| Uses timing of the event? | Yes — central to the model | No — ignores exactly when it happened |
| Handles participants lost to follow-up? | Yes, via censoring (see below) | Not directly — typically excluded or treated as no-event |
| Effect measure | Hazard ratio (HR) | Odds ratio (OR) |
| Typical question | "How much faster/slower does the event happen?" | "How much more/less likely is the event by a fixed point?" |
What Is Survival Analysis?
Survival analysis is the broader family of statistical methods, including Cox regression, designed specifically for time-to-event data. "Survival" is the traditional name from its origins in mortality research, but the methods apply equally to any event, not only death — recurrence, relapse, discharge, or infection all count.
The two most common survival analysis tools work together: Kaplan-Meier estimates and visualizes survival probability over time for one or more groups without adjusting for other variables, while Cox regression quantifies the effect of one or more predictors on survival time, with or without adjustment for other factors. Most survival analysis papers present both: a Kaplan-Meier curve to show the overall picture, and a Cox model to quantify and adjust the effect. See our full Kaplan-Meier guide for the visual half of this pairing.
What Is Censoring?
Censoring happens when a study ends, or a participant leaves the study, before the event of interest has occurred for them — meaning their exact event time is unknown, but you do know they survived event-free for at least as long as they were observed. This is the single most important concept that makes survival analysis different from ordinary regression, because it lets a study use every participant's partial information rather than discarding anyone whose event hadn't happened by the study's end.
Patient A enrolls and dies at Month 14 — event observed, not censored.
Patient B is still alive when the 24-month study ends — censored at Month 24 (the study ended before their event, if any, occurred).
Patient C moves away and is lost to follow-up at Month 9, alive at last contact — censored at Month 9 (their event status after Month 9 is unknown).
All three types shown above are examples of right censoring, by far the most common type in medical research (the true event time is somewhere to the right of, i.e., after, the last known observation). Cox regression correctly uses censored participants' data up to the point they were censored — it does not simply drop them, and it does not treat censoring as equivalent to "no event ever."
Hazard vs Risk
"Hazard" and "risk" are often used interchangeably in everyday language, but in survival analysis they mean two distinct things, and confusing them is a common source of misinterpretation.
| Concept | What It Measures | Simple Analogy |
|---|---|---|
| Risk | The cumulative probability that an event has happened by a specific point in time (a fixed number between 0 and 1) | "What fraction of the road trip is behind you by mile 100?" |
| Hazard | The instantaneous rate of the event happening right now, among those who haven't yet had it — can change at every moment | "How fast is the car going right now, at this exact instant?" |
A hazard is not a probability and does not have to stay between 0 and 1 — it is a rate, similar to speed. A hazard ratio compares this instantaneous rate between two groups, at any given moment, assuming that ratio stays roughly constant over the whole follow-up period (this assumption is covered in detail later in this guide).
Hazard Ratio Explained with Simple Examples
The hazard ratio (HR) is the single number Cox regression is built to produce. It compares the hazard (instantaneous event rate) in one group to the hazard in a reference group. An HR of 1.0 means no difference; above 1.0 means a higher hazard (faster/more frequent events); below 1.0 means a lower hazard (slower/less frequent events, i.e., protective).
HR = 1.8 for smokers vs non-smokers (outcome: lung cancer diagnosis) — at any given moment during follow-up, smokers who have not yet been diagnosed are experiencing new diagnoses at 1.8 times the rate of non-smokers who have not yet been diagnosed. This is not the same as saying "smokers have an 80% higher chance of getting lung cancer by the end of the study" — that would be a statement about cumulative risk, not hazard.
"HR = 1.8 means smokers have an 80% higher probability of developing lung cancer over their lifetime" — treating the hazard ratio as if it were a cumulative risk or probability statement.
"At any point during follow-up, smokers who have not yet developed lung cancer are being diagnosed at 1.8 times the instantaneous rate of non-smokers who have not yet been diagnosed."
Adjusted vs Unadjusted Hazard Ratios
An unadjusted (crude) hazard ratio comes from a Cox model containing only the exposure or treatment variable, with nothing else — it reflects the raw, unadjusted relationship, which may be distorted by confounding variables. An adjusted hazard ratio comes from a model that also includes other covariates (age, disease stage, comorbidities), isolating the exposure's effect from their influence. See our guide on how to choose covariates for the full logic behind deciding which variables belong in the adjustment set.
Unadjusted: New drug vs standard care, HR = 0.72 (28% lower hazard of progression) — but patients on the new drug happened to have earlier-stage disease on average.
Adjusted (for tumor stage and age): HR = 0.81 — still protective, but part of the original crude effect was explained by the imbalance in disease stage between groups, not the drug itself.
Always report which covariates a hazard ratio was adjusted for, and generally report both the unadjusted and adjusted HR side by side so a reader can see how much confounding adjustment changed the estimate.
Interpreting the SPSS Cox Regression Output Table
Run via Analyze → Survival → Cox Regression in SPSS. The key output table is Variables in the Equation, shown below with realistic values for a study of a new drug's effect on disease progression, adjusted for age and tumor stage.
What to Interpret, What to Ignore
Interpret Exp(B) — this is the hazard ratio itself, the number to report and discuss. Ignore the raw B (log-hazard) column in your write-up; it is a computational intermediate, not a clinically meaningful number on its own — similar to how logistic regression's raw B is not reported directly either, covered in our SPSS output interpretation guide. The Wald statistic is the test statistic behind the p-value; it rarely needs to be quoted directly, but the p-value it produces does.
Interpreting Confidence Intervals for the Hazard Ratio
The 95% confidence interval (CI) around a hazard ratio gives the plausible range for the true hazard ratio in the population, based on your sample. A CI that does not cross 1.0 confirms the result is statistically significant at the conventional threshold — consistent with, but conveying much more than, the p-value alone.
| Example HR (95% CI) | Interpretation |
|---|---|
| 0.58 (0.39–0.85) | Significant protective effect — entire interval is below 1.0 |
| 1.45 (1.05–2.01) | Significant harmful effect — entire interval is above 1.0 |
| 0.85 (0.60–1.20) | Not significant — interval crosses 1.0; true effect could be protective, null, or harmful |
| 0.58 (0.09–3.71) | Extremely imprecise — likely too few events to draw a reliable conclusion, despite a similar point estimate to the first row |
Always report the full CI, not just the point estimate — a "significant" HR with a very wide CI (like the last row above) deserves far more caution than the same HR with a tight CI, even though both might report similar p-values. See our confidence interval guide for the general concept behind this.
Interpreting P Values in Cox Regression
The p-value in the Sig. column tests the null hypothesis that the true hazard ratio equals 1.0 (no effect). A p-value below your chosen threshold (almost always 0.05) means the observed HR is unlikely to have arisen by chance alone if the true HR were really 1.0.
As with any regression, always report the exact p-value (e.g., p = 0.007), not just "p < 0.05," and convert any SPSS ".000" to "p < 0.001" rather than reporting it literally as zero. A p-value alone says nothing about the size of the effect — always pair it with the HR and its CI, exactly as covered in our p-value interpretation guide.
Checking the Proportional Hazards Assumption
Cox regression assumes the hazard ratio between groups stays roughly constant over the entire follow-up period — this is the "proportional" in "proportional hazards." If a treatment's benefit is large early on but fades by month 24, a single overall hazard ratio poorly summarizes what is actually a changing effect, and the model's p-value and CI become unreliable.
Visual check: log-minus-log survival plot
Plot log(-log(survival)) against log(time) for each group. Roughly parallel lines support the assumption; lines that cross or diverge suggest violation.
Formal test: Schoenfeld residuals
Test each covariate's Schoenfeld residuals against time. A non-significant result (p > 0.05) supports proportional hazards for that variable; a significant result suggests violation.
If violated: adapt the model
Options include a time-dependent covariate, stratifying by the violating variable, splitting follow-up into time periods, or an alternative model that doesn't require this assumption.
Common Mistakes in Interpreting Cox Regression
Mistake 1: Reading the Hazard Ratio as a Risk or Probability
Saying "HR = 1.8 means an 80% higher chance of the event" converts a rate comparison into a probability statement, which is not what a hazard ratio measures.
Mistake 2: Reporting the Raw B Instead of Exp(B)
The log-hazard coefficient (B) is not directly interpretable by a clinical reader and should never be reported as if it were the hazard ratio.
Mistake 3: Excluding Censored Participants from the Analysis
Dropping everyone who didn't experience the event by the end of the study discards real information and can bias the result, sometimes substantially.
Mistake 4: Never Checking the Proportional Hazards Assumption
Reporting a single hazard ratio without checking whether the effect is actually constant over time can mask a clinically important pattern, such as an early benefit that disappears later.
Mistake 5: Confusing Hazard Ratio with Odds Ratio or Relative Risk
These three effect measures are calculated differently and are not numerically interchangeable, even though they often point in a similar direction.
Mistake 6: Claiming Causation from an Observational Cox Model
An adjusted hazard ratio from an observational study, no matter how many covariates were included, cannot rule out unmeasured confounding the way randomization can.
Step-by-Step Medical Example
Study: Does a new anticoagulant reduce the hazard of stroke in patients with atrial fibrillation, compared to standard therapy, over 24 months of follow-up?
Define the outcome
Event = ischemic stroke (yes/no); Time = months from enrollment to stroke, or to last contact/study end if stroke-free (censored).
Choose covariates
Age, prior stroke/TIA, hypertension, and diabetes — all established stroke risk factors, selected via clinical knowledge and the CHA₂DS₂-VASc score, as covered in our covariate selection guide.
Run the Cox model
Analyze → Survival → Cox Regression in SPSS; treatment group as the primary predictor, the four covariates entered together.
Check the proportional hazards assumption
Schoenfeld residuals test for treatment group: p = 0.41 (non-significant) — assumption holds.
Interpret and report the result
Adjusted HR for the new anticoagulant = 0.62, 95% CI [0.41, 0.94], p = 0.024.
"After adjusting for age, prior stroke/TIA, hypertension, and diabetes, patients receiving the new anticoagulant had a significantly lower hazard of ischemic stroke compared to standard therapy (aHR 0.62, 95% CI [0.41, 0.94], p = 0.024). The proportional hazards assumption was confirmed via Schoenfeld residuals (p = 0.41)." Run this analysis directly with the Cox regression calculator, and see our full reporting-by-test guide for more worked examples across other tests.
Practical Interpretation Checklist
Confirm event and time variables are correctly defined
Event coded consistently (1 = event, 0 = censored); time measured from a clear, consistent starting point.
Report Exp(B), not raw B
The hazard ratio is the exponentiated coefficient — always report this, never the log-hazard.
Report the 95% CI alongside every HR
Never quote a hazard ratio without its confidence interval.
Report the exact p-value
Three decimal places, or "p < 0.001" — never "p = 0.000" or bare "p < 0.05."
State whether the HR is adjusted or unadjusted
And list every covariate included if adjusted.
Check and report the proportional hazards assumption
Schoenfeld residuals test or log-minus-log plot, for the primary predictor at minimum.
Use rate language, not probability language
"X times the instantaneous hazard," not "X% more likely."
Match causal language to study design
Association language for observational studies; causal language reserved for randomized designs.
Frequently Asked Questions
Run Your Cox Regression With Confidence
Use StatClinic's free Cox regression and Kaplan-Meier calculators to run your survival analysis and get a journal-ready, correctly worded results sentence generated automatically. Free, no registration required.
Try StatClinic Free →