A Primer on Time-to-event Analysis

Meta-analysis Academy

Time-to-event analysis is employed to assess the prognostic influence of an exposure

or intervention in situations where time is a crucial factor. This approach goes beyond merely

identifying which group experiences a higher outcome incidence during a specified follow-

up, as it also assesses the rate at which the incidence progresses over time. Conventional

time-to-event analysis is centered on comparing the duration between the initiation of

exposure or intervention and the occurrence of a relevant event.

We now delve into comprehending the hazard, a pivotal concept in traditional time-

to-event analysis methods. Please notice how the hazard encapsulates information regarding

both outcome risk and timing into a singular measure. In the time-to-event analysis of

exposure or intervention effects, comparisons between groups are drawn based on hazard

ratios.

Hazard

In simpler terms, the hazard at a specific time denotes the probability of experiencing

the event of interest in the upcoming interval among individuals who had not yet experienced

the event by the beginning of that interval.

It is important to observe that the hazard was defined "at any given time" rather than

"for every follow-up time." This signifies that the hazard is a function of time, h(t), and can

be illustrated as a continuous curve, depicting how the "instantaneous" risk evolves over

time.

The hazard is quantified as a rate of events per unit of time. However, it is not merely

derived by dividing the number or risk of events by the follow-up time. Calculating the
hazard function poses a challenge because the event of interest is not observed in all patients.

This implies that, for certain patients, the occurrence of the event remains invisible — or

censored. For these individuals, determining the exact timing of the event becomes

impossible.

The information within censored observations varies. Patients with early censored

time-to-event data (e.g., those lost to follow-up in the initial weeks of a study) contribute less

information than those observed for an extended period before censoring. Nevertheless, every

observation contributes some information. Therefore, time-to-event analysis methods must

accommodate censoring. In this tutorial, we will exclusively focus on the most commonly

used methods: Kaplan-Meier curves and Cox proportional hazards models.

Kaplan-Meier Curves

A Kaplan-Meier curve illustrates the proportion of "surviving" patients (those who

have not experienced an outcome) over time for each treatment group. The term "survival" is

employed because time-to-event analysis, more conventionally referred to as "survival

analysis," was initially developed to assess mortality.

The height of the Kaplan-Meier curve at the conclusion of each time interval is

determined by multiplying the fraction of patients who remained event-free at the end of the

preceding interval by the fraction of patients who survive the current time interval without

experiencing an event. This calculation yields the proportion of patients expected to remain

event-free at the end of the current interval. The value of the Kaplan-Meier curve at the

current time interval then becomes the starting value for the next time interval.

The cumulative multiplication process initiates with the first time interval (study start)

and progresses stepwise throughout the entire follow-up. Hence, Kaplan-Meier rates are also

termed the "product-time estimate" of the survival curve. This method is designed to
appropriately address censoring, as it exclusively considers patients still under observation at

the commencement of each time interval when determining the fraction "surviving" at the

interval's conclusion.

The cumulative incidence of events can be calculated as 1 minus the Kaplan-Meier

rates. Plots of cumulative incidences essentially mirror the traditional Kaplan-Meier plot in

an "upside-down" fashion, offering similar information.

Log-rank tests

Although a Kaplan-Meier plot visually displays variations in survival curves between

groups over time, it does not provide a direct indication of their statistical significance. The

assessment of statistical significance for observed differences is conducted through the log-

rank test. It is essential to note that this test does not have the capacity to control for

confounding variables, such as disparities in patient demographics between groups.

Cox proportional hazards model

The Cox proportional hazards model not only tackles the issue of censoring but also

enables adjustment for numerous prognostic independent variables and confounders, such as

age and sex. The model operates on the assumption that there is a "baseline" hazard function

for individuals with independent predictor variables (such as exposure or intervention) equal

to their reference (control) value. Notice that the baseline function is not explicitly defined

and is allowed to take any shape.

The Cox model yields a hazard ratio for each independent predictor variable. The

hazard ratio quantifies the extent to which the "baseline" hazard is multiplied for every unit

change in the variable of interest. In the case of binary independent variables, where the

control group has a value of 0 and the group of interest has a value of 1 for the variable
defining exposure or intervention, the hazard ratio provides insight into the proportional

change in hazard associated with the variable.

Hazard ratios share a numerical interpretation similar to risk ratios: values greater

than 1 signify increased risk in the group of interest, values less than 1 indicate decreased

risk, and values of 1 suggest comparable risk between groups. However, as defined

previously, hazards are not cumulative risks. Therefore hazard ratios do not have the same

clinical interpretation as risk ratios.

Caveats

The Cox proportional hazards model relies on two key assumptions. First, censoring

must be independent of outcome occurrence. If the probability of censoring is correlated with

the risk of the event, the hazard ratio becomes entirely invalid, as it ends up measuring the

impact of this bias rather than assessing the true effect of an exposure or intervention. To

illustrate, consider a clinical trial studying the impact of a new anti-obesity medication on

cardiovascular outcomes. An example of this bias would be if patients in the control group

were less likely to follow up with trial investigators because they did not experience the

expected weight loss.

Second, hazard functions should be proportional among different patient groups.

Essentially, this means that the hazard functions should share the same shape, differing only

in magnitude. While not meeting this assumption does not invalidate the overall

interpretation of the hazard ratio, it does restrict the conclusion that the observed hazard ratio

applies uniformly across all follow-up times. The accurate interpretation is that hazard ratios

vary over time, and the observed hazard ratio represents a time-weighted average of the time-

varying hazard ratios. In practice, this assumption is often not met in medicine, as it is
uncommon for an exposure or intervention to exert a consistent effect over time. Testing this

assumption is typically done by examining Schoenfeld residuals, though all tests designed for

this purpose are frequently underpowered.

Practical questions

1. Using the Kaplan-Meier curve for the primary endpoint of the FOURIER trial (DOI:

10.1056/NEJMoa1615664), reconstruct individual patient data for each study group. (See the

video tutorials available in the “User Guide” section of this website:

[Link]

2. Using the tidyverse ([Link] (1) Merge patient data derived from

Kaplan-Meier curves of FOURIER treatment arms; (2) Simplify variable names for

consistency (put everything in lower case, separated by underscores); (3) view the final,

wrangled data.

3. Using the survival ([Link] and

ggsurvfit ([Link] packages, create a

Kaplan-Meier curve.

4. Using the gtsummary R package ([Link] obtain the

hazard ratio and log-rank p values.

Harrell FE. Regression modeling strategies: With applications to linear models, logistic and

ordinal regression, and survival analysis. 2nd ed. Cham, Switzerland: Springer International

Publishing; 2015.

Tolles J, Lewis RJ. Time-to-Event Analysis. In: Livingston EH, Lewis RJ. eds. JAMA Guide

to Statistics and Methods. McGraw-Hill Education; 2019. Accessed February 16, 2024.

[Link]

Stensrud MJ, Hernán MA. Why Test for Proportional Hazards? JAMA. 2020;323(14):1401–

1402. doi:10.1001/jama.2020.1267

Hernán MA. The hazards of hazard ratios. Epidemiology. 2010 Jan;21(1):13-5. doi:

10.1097/EDE.0b013e3181c1ea43. Erratum in: Epidemiology. 2011 Jan;22(1):134. PMID:

20010207; PMCID: PMC3653612.

Han L. Breaking Free from the Hazard Ratio: Embracing the Restricted Mean Survival Time

in Clinical Trials. NEJM Evid. 2023 Aug;2(8):EVIDe2300142. doi: 10.1056/EVIDe2300142.

Epub 2023 Jul 25. PMID: 38320152.