STATISTICS AND RESEARCH DESIGN

Survival analysis, part 3: Cox regression

Despina Koletsia and Nikolaos Pandisb
London, United Kingdom, Athens and Corfu, Greece, and Bern, Switzerland

A
nother approach to analyze survival data is to
Table I. Cox regression for the effect of wire on
use regression analysis. This can be accom-
instant probability of alignment
plished by applying Cox regression, which allows
us to calculate a special form of rate ratios known as Hazard ratio 95% CI P value
hazard ratios.1 Therefore, instead of just a P value pro- Type of wire
vided by the log-rank test, we can obtain an effect esti- Wire A Reference
mate with its confidence intervals, and at the same time Wire B 1.28 0.70-2.36 0.42
we can adjust for other categorical or continuous cova-
riates and interactions.
Table II. Cox regression for the effect of wire on prob-
Hazard vs rate ability of alignment after adjusting for crowding
The true distinction is that a hazard is instantaneous, Hazard ratio 95% CI P value
whereas a rate is calculated over a period of time (and Type of wire
usually it is implicitly assumed to be constant over this Wire A Reference
period).2 Wire B 1.43 0.78-2.65 0.25
An analogy could be made with speed; the rate is like Crowding Per unit (mm)
average speed: you take the total distance you drove and 0.73 0.60-0.88 0.001
divide it by the amount of time it took. The hazard is
take it into account, we might incorrectly infer that 1
what your speedometer showed at any point in time.
wire is inferior or superior to the other.
In Table I, we used Cox regression to perform the
same analysis as in the previous article with the log-
rank test.3 Interpretation
The hazard ratio of 1.28 indicates that the hazard The hazard (instant probability of reaching align-
(instant probability) to experience the event (reach align- ment) is 1.44, and it indicates that wire B has a 44%
ment) is 28% higher for wire B compared with wire A. higher hazard to reach alignment compared with wire
The interpretation is similar to the previously encoun- A after adjusting for pretreatment crowding (95% CI,
tered regression output.4 The P value is 0.423 (same as 0.78-2.65; P 5 0.25). The pretreatment amount of
the log-rank test), and it indicates no statistically signif- crowding is a significant predictor, and it indicates
icant difference between the wires in terms of the instant that for a 1-mm increase in pretreatment crowding,
probability of reaching the alignment. there is a 27% (27 5 1 – 0.73) decrease in the instant
In Table II, we compare the rates of alignment be- probability of reaching alignment.
tween wire B vs wire A after adjusting for the amount
of crowding. This is logical, since it is likely that in small Cox regression vs logistic regression vs linear
trials there could be imbalances between treatment regression
groups in terms of crowding. If 1 group has greater
mean crowding, then it is likely that the time to reach Although using survival analysis is different from us-
alignment (the event) will increase; hence, if we do not ing logistic regression, the conclusions are similar.
As a reminder and to help make the connection with
a
London School of Hygiene and Tropical Medicine, University of London, Lon- logistic regression, we ask “what are the odds of reaching
don, United Kingdom; private practice, Athens, Greece. alignment in the wire B group compared with the wire A
b
Department of Orthodontics and Dentofacial Orthopedics, School of Dental
Medicine/Medical Faculty, University of Bern, Bern, Switzerland; private practice, group without accounting for time elapsed (expressed as
Corfu, Greece. the odds ratio)?”
Am J Orthod Dentofacial Orthop 2017;152:722-3 In survival analysis, we are asking “what is the instant
0889-5406/$36.00
Ó 2017. probability of reaching alignment in the wire B group vs
http://dx.doi.org/10.1016/j.ajodo.2017.07.009 the wire A group (hazard ratio)?” This model includes the
722
Statistics and research design
Table III. Comparison of linear, logistic, and Cox regression
Method Outcome predictor
Linear regression Continuous
Logistic regression Binary
Cox regression Binary
Fig. Nelson-Aalen plot of cumulative hazard by wire type
element of time, whereas logistic regression does not.
Table III compares some aspects of Cox, logistic, and
linear regression.
More on Cox regression and the proportional
hazards assumption
A Cox model is based on very small intervals of time,
called time-clicks, which contain at most 1 event. This
produces a constant rate within each small interval, but
also it allows the rate to vary over longer time intervals.
For this reason, Cox regression is used to model rates
that change very rapidly. In this situation, the numerator
of the rate is at most 1, and the denominator is the number
of subjects at risk when the event occurs, which is called
the risk set. Therefore, risk sets are matched on time.
The Cox survival is valid when the proportional haz-
ard assumption is satisfied. This implies that the hazard
can vary with time, whereas the hazard ratio (the effect
measure—ie, wire B vs wire A—or the ratio of the rates
of exposed to the baseline category of nonexposed sub-
jects in observational studies) must stay constant. This
can be verified visually with various graphs and appro-
priate tests.
There is a plot of the cumulative hazard (on a log
scale against time), known as the Nelson-Aalen, in
which the proportional hazard assumption is not
violated only if the lines tend to be parallel, as shown
in the Figure.
American Journal of Orthodontics and Dentofacial Orthopedics
723
Predictor outcome Censoring allowed
Continuous and/or categorical No
Continuous and/or categorical No
Continuous and/or categorical Yes
If the proportional hazard assumption is violated, the
data can be partitioned into periods in which the hazard
ratio remains constant; hence, separate models can be
fitted for the partitioned periods, data can be trans-
formed, or parametric survival models can be used.
Cox regression allows the baseline rate to vary, but it
assumes that the rate ratio remains proportional
throughout the follow-up period. More specifically,
Cox regression is termed “proportional hazards regres-
sion.”
Overall, Cox proportional hazard regression is a flex-
ible tool that models the hazard, and it describes the
probability of failure during a very small time increment,
assuming that no failures have occurred before that
time. The baseline hazard is not specified, and the haz-
ards of the covariates are assumed to be proportional to
the time of the baseline hazard. The results are reported
as a hazard ratio, which can be interpreted as the relative
risk of an event for each covariate.
REFERENCES
1. Breslow NE. Analysis of survival data under the proportional hazards
model. Int Stat Rev 1975;43:45-57.
2. Kirkwood BR, Sterne JAC. Essential medical statistics. 2nd ed. Ox-
ford, United Kingdom: Blackwell Science; 2003. p. 287-94.
3. Koletsi D, Pandis N. Survival analysis, part 2: Kaplan-Meier method and
the log-rank test. Am J Orthod Dentofacial Orthop 2017;152:569-71.
4. Pandis N. Linear regression. Am J Orthod Dentofacial Orthop 2016;
149:431-4.
November 2017  Vol 152  Issue 5