Incidence Rate, Hazard, and Cox Regression: Distinctions in Time-to-Event Research

Table of Different slices of risk in time-to-event research:

Concept	Core question answered	Typical notation (plain text)	Where is it used
Incidence Rate	“How many events occur per unit of person-time?”	Rate = events ÷ person-time	Descriptive summaries; Poisson or negative-binomial regression
Hazard (instantaneous rate)	“Given someone is still event-free right now, what is the chance the event strikes in the next instant?”	h(t)	Survival curves; parametric survival models; Cox regression
Cox (hazard) regression	“How does the hazard change when we alter a covariate, without forcing a specific hazard shape over time?”	h(t\|x) = h₀(t) × exp(β × x)	Adjusted survival analyses; hazard-ratio modelling

1 Incidence Rate (“Person-time rate”)

What it is

Counts how quickly events accumulate across everyone’s observed follow-up.

Incidence rate=Number of eventsTotal observed person-time\text{Incidence rate} = \frac{\text{Number of events}}{\text{Total observed person-time}}Incidence rate=Total observed person-timeNumber of events

How to picture it

Imagine ten asthma patients followed for one year each (total = 10 patient-years). If three attacks occur, the rate is 0.3 attacks per patient-year. It treats follow-up like fuel consumed: every month contributes to the denominator until an attack or censoring.

When it shines

Comparing disease burden across populations of different sizes or follow-up lengths.
Fitting Poisson or negative-binomial regression when events are rare and hazards are roughly constant within chosen intervals.

2 Hazard (“Instantaneous rate”)

What it is

A microscope view of risk. For any time t, the hazard h(t) is the probability that an event occurs in the next infinitesimal moment given survival up to t, divided by that tiny time width.

Hazard Function Definition:

h (t) = lim_{Δt \to 0} \frac{P (t \leq T < t + Δt | T \geq t)}{Δt}

Variable Definitions:

h(t): Hazard function at time t, the instantaneous rate of occurrence of the event at time t, given survival until time t.
t: A specific point in time at which the hazard is evaluated.
T: A random variable representing the time until the event of interest (e.g., failure or death).
Δt: A small time interval approaching zero.
P(t ≤ T < t + Δt | T ≥ t): Conditional probability that the event occurs in the interval [t, t + Δt) given that it has not occurred before t.

Why does it differ from the incidence rate

The incidence rate looks backward over an entire follow-up interval.
The hazard looks forward at an instant, continuously updating as the risk set shrinks and as biology or treatment effects evolve.

Typical shapes

Constant → exponential survival (e.g., radioactive decay).
Rising → ageing mortality.
Peaking then falling → postoperative complications are highest in week 1, then taper.

3 Cox Proportional-Hazards Regression (“Hazard regression”)

Key idea

Cox regression models relative changes in the hazard without specifying its baseline shape:

log h(t | x) = log h0(t) + β1 x1 + β2 x2 + …

h0(t) = baseline hazard (left completely free-form).
e^βk= Hazard Ratio (HR) for a one-unit increase in xk at every instant—the proportional-hazards assumption.

What the model gives you

Adjusted HRs (“instantaneous risk multipliers”) similar to risk ratios but varying over time only through h0(t), not through covariates.
No absolute survival times unless you also estimate or predict the baseline curve.

When to use

You need relative effects of treatments or prognostic factors and are unsure of the baseline hazard’s shape.
The proportional-hazards assumption is reasonable or can be repaired (e.g., stratification, time-interaction terms).

Putting it all together

Question you’re asking	Best metric/tool
“How common are events over 1,000 patient-days?”	Incidence rate
“Is the moment-to-moment risk climbing or fading?”	Hazard curve (parametric or smooth estimate)
“Does Treatment A cut that moment-to-moment risk in half compared with Treatment B, after age adjustment?”	Cox regression HR

Worked micro-example

Study: 200 post-surgical patients followed to wound infection or 30 days.Results: 15 infections in 5,000 patient-days.

Incidence rate = 15 ⁄ 5,000 = 0.003 infections/day.
Hazard at day 3 estimated (say, via kernel smoothing) = 0.008/day—higher than the average because early days are riskier.
Cox model, adjusting for antibiotic prophylaxis and diabetes, shows Antibiotics HR = 0.55 → patients on antibiotics have 45 % lower instantaneous risk at every postoperative moment (assuming PH).

Quick rule-of-thumb distinctions

Rate → events ÷ time (whole follow-up), useful for burden and Poisson models.
Hazard → instantaneous risk, forms the backbone of survival curves.
Cox regression → log-hazard model producing adjusted hazard ratios when the baseline shape is unknown or unimportant.