Poisson and Binary Regression

Janne Pitkäniemi
Finnish Cancer Registry
Tampere university

Statistical Practice in Epidemiology (2024,Lyon)

1 / 20
Elapse of time and Epidemiology
Epidemiology deals with the occurence of event (disease) in populations observed
over time
▶ concepts of risk and rate are used to measure the frequency with which the
event (disease) cases occur
▶ risk is dened as D N , where D is the number of people who developed the
disease during pre-specied follow-up from 0 to t and N is the number of
disease-free population at the beginning of follow-up and
▶ rate is dened as YD , where Y is the amount of person-time at risk observed
when following disease free subjects from 0 to t.
▶ Note: risk increases with t but rate can vary depending on the length of the
follow-up period.
▶ Virtually all prospective follow-up studies include loss to follow-up
censoring and risk must be estimated using appropriate methods
described in this course.
2 / 20
Points to be covered
▶ Incidence rates, rate ratios and rate dierences from
follow-up studies can be computed by tting Poisson regression models.

▶ Risk ratios and dierences can be computed from binary data by tting
Logistic regression models.

▶ Both models are special instances of

Generalized linear models.

▶ There are various ways to do these tasks in R.

3 / 20
The Estonian Biobank cohort: survival among the elderly
Follow-up of 60 random individuals aged 75-103 at recruitment, until death (•)
or censoring (o) in April 2014 (linkage with the Estonian Causes of Death
Registry). (time-scale: calendar time).
60

●
●
● ●
● ●
● ●
●
50

●
●
● ●
● ●
● ● ●
●
40

● ●
● ●
●
●
●
index

● ●
●
30

●
●
● ●
● ●
●
●
●●
20

●
●
●
●
●
●
● ● ●
●
10

●
●
●
●
● ●
● ●
● ●
●
0

2004 2006 2008 2010 2012 2014

Time

4 / 20
The Estonian Biobank cohort: survival among the elderly
Follow-up time for 60 random individuals aged 75-103 at recruitment (time-scale:
time in study).
60

● ●
● ●
● ●
● ●
● ●
50

● ●
●
●
●
● ● ●
●
40

● ●
● ●
●●
●
index

● ●
●
30

●●
● ●
● ● ●
●
●●
20

●
●
●●
●
●
● ● ●
●
10

●
●● ●●
● ●
● ●
● ●
0

0 2 4 6 8

Time (years since rectuitment)

5 / 20
Events, dates and risk time
▶ Mortality as the outcome:
d: indicator for status at exit:
1: death observed
0: censored alive
▶ Dates:

doe = date of Entry to follow-up,

dox = date of eXit, end of follow-up.

▶ Follow-up time (years) computed as:

y = (dox - doe)/365.25

6 / 20
Crude overall rate computed by hand and model
Total no. cases, person-years & rate (/1000 y):
> D <= sum( d ); Y <= sum(y) ; R <= D/(Y/1000)
> round( c(D=D, Y=Y, R=R), 2)
D Y R
884.00 11678.24 75.70
R-implementation of the rate estimation with Poisson regression:
A model with oset term A model with poisreg=family (Epi package)
> m1 <= glm( D ~ 1, family=poisson,
oset=log(Y)) > glm(cbind(D, Y) ~1, family=poisreg)

> coef(m1) Coecients :

( Intercept ) ( Intercept )
=2.581 =2.581
From the coecient we get estimate of the rate exp(−2.581) ∗ 1000 = 75.70
7 / 20
Constant hazard  Poisson model
Let Y ∼ exp (λ), then f (y ; λ) = λe −λy I (y > 0)
Constant rate model: λ(y ) = Sf ((yy ;λ)
;λ)
= λ and observed data {(yi , δi ); i = 1, ..., n}.

The likelihood L(λ) = ni=1 λδ e −λy and

Q
i i

n
[δi log (λ) − λyi ]
P
log (L) =
i =1
Solving the score equations:
∂ log L(λ)
− yi = Dλ − Y = 0 and − λY = 0
P δ 
∂λ
= λ
i
D

→ maximum likelihood estimator (MLE) of λ:

D number of cases
λ
b= = = empirical rate!
Y total person-time
8 / 20
oset term  Poisson model
▶ Previous model without oset: Intercept 6.784=log(884)
▶ We should use an oset if we suspect that the underlying population sizes
(person-years) dier for each of the observed counts  For example
varying person-years by sex,age,treatment group,...
▶ We need a term in the model that "scales" the likelihood, but does not
depend on model parameters ( include a term with reg. coef. xed to
1)  oset term is log(y)
▶ This is all taken care of by family=poisreg  recommend to use

y ) = β0 + β1 x1
µ
log (

log (µ) = 1 × log (y ) + β0 + β1 x1

9 / 20
Comparing rates: The Thorotrast Study
▶ Cohort of seriously ill patients in Denmark on whom angiography of brain
was performed.
▶ Exposure: contrast medium used in angiography,
1. thor = thorotrast (with 232Th), used 1935-50
2. ctrl = other medium (?), used 1946-63
▶ Outcome of interest: death

doe = date of Entry to follow-up,

dox = date of eXit, end of follow-up.

▶ data(thoro) in the Epi package.

10 / 20
Tabulating rates: thorotrast vs. control
Tabulating cases, person-years & rates by group
> stat. table ( contrast ,
+ list ( N = count(),
+ D = sum(d),
+ Y = sum(y),
+ rate = ratio(d,y,1000) ) )
============================================
contrast N D Y rate
============================================
ctrl 1236 797.00 30517.56 26.12
thor 807 748.00 19243.85 38.87
============================================

11 / 20
Rate ratio estimation with Poisson regression
▶ Include contrast as the explanatory variable (factor).
▶ Insert person years in units that you want rates in
> m2 <= glm( cbind(d,y/1000) ~ contrast,family = poisreg(link="log") )
> round( summary(m2)$coef, 4)[, 1:2]

Estimate Std. Error

( Intercept ) 3.2626 0.0354
contrast thor 0.3977 0.0509
▶ Rate ratio and CI?
Call function ci.exp() in Epi
> round( ci.exp( m2 ), 3 )
exp(Est.) 2.5% 97.5%
(Intercept) 26.116 24.364 27.994
contrast thor 1.488 1.347 1.644
12 / 20
Rates in groups with Poisson regression
▶ Include contrast as the explanatory variable (factor).
▶ Remove the intercept (-1)
▶ Insert person-years in units that you want rates in
> m3 <- glm( cbind(d,y/1000) ~ factor(contrast)-1,family = poisreg)
> round( summary(m3)$coef, 4)[, 1:2]

Estimate Std. Error

contrast ctrl 3.2626 0.0354
contrast thor 3.6602 0.0366

> round( ci.exp( m3 ), 3 )

exp(Est.) 2.5% 97.5%

contrast ctrl 26.116 24.364 27.994
contrast thor 38.870 36.181 41.757

13 / 20
Rate dierence estimation with Poisson regression
▶ The approach with d/y enables additive rate models too:
> contrast<-c(0,1)
> m5 <-glm(cbind(d,y/1000) ~contrast,
family=poisreg(link="identity") )
> round( ci.exp(m5,Exp=F), 3 )

Estimate 2.5% 97.5%

(Intercept) 26.116 24.303 27.929
contrast thor 12.753 9.430 16.077

14 / 20
Binary data: Treatment success Y/N
85 diabetes-patients with foot-wounds:
▶ Dalterapin (Dal)
▶ Placebo (Pl)
Treatment/Placebo given to diabetes patients, the design is prospective and
outcome is measured better(Y)/worse(N). Is the probability of outcome more
than 15%  yes, then use the risk dierence or risk ratio (RR)
Treatment group
Dalterapin Placebo
Better 29 20
Worse 14 22
Total 43 42
29 20
p̂ Dal = 43 = 67% p̂ Pl = 42 = 47%
15 / 20
Binary data: Crosstabulation analysis of 2x2 table
> library(Epi)
> dlt <- rbind( c(29,14), c(20,22) )
> colnames( dlt ) <- c("Better","Worse")
> rownames( dlt ) <- c("Dal","Pl")
> kable(twoby2( dlt ),"latex")

2 by 2 table analysis:
Better Worse P(Better) 95% conf. interval
Dal 29 14 0.6744 0.5226 0.7967
Pl 20 22 0.4762 0.3316 0.6249
95% conf. interval
Relative Risk: 1.4163 0.9694 2.0692
Sample Odds Ratio: 2.2786 0.9456 5.4907
Conditional MLE Odds Ratio: 2.2560 0.8675 6.0405
Probability difference: 0.1982 -0.0110 0.3850

Exact P-value: 0.0808

Asymptotic P-value: 0.0665

16 / 20
Binary regression  estimation of odds ratio
For grouped binary data, the response is a two-column matrix with columns
(successes,failures).
> library(Epi)
> library(xtable)
> dlt <- data.frame(rbind( c(29,14),c(20,22) ))
> colnames( dlt ) <- c("Better","Worse")
> dlt$trt <- c(1,0)
> b2<-glm(cbind(Better,Worse)~trt,
+ family=binomial(link="logit"),
+ data=dlt)
> xtable(round( ci.exp( b2 ), digits=6 ))
exp(Est.) 2.5% 97.5%
(Intercept) 0.91 0.50 1.67
trt 2.28 0.95 5.49
▶ The default parameters in logistic regression are odds (the intercept:
20/22 = 0.9090) and the odds-ratio ((29/14)/(20/22) = 2.28).
▶ This is NOT what you want, because odds ratio is biased estimate of the
risk ratio.(recall if p>10% 1−p p ̸≈ p )
17 / 20
Binary regression - Estimation of risk ratio (Relative risk)
> library(Epi)
> library(xtable)
> dlt <- data.frame(rbind( c(29,14),c(20,22) ))
> colnames( dlt ) <- c("Better","Worse")
> dlt$trt <- c(1,0)
> b2<-glm(cbind(Better,Worse)~trt,
+ family=binomial(link="log"),
+ data=dlt)
> xtable(round( ci.exp( b2 ), digits=6 ))
exp(Est.) 2.5% 97.5%
(Intercept) 0.48 0.35 0.65
trt 1.42 0.97 2.07
Diabetics with Dalterapin treatment are 1.4 times likely to get better than those
treated with placebo
18 / 20
Binary regression - Estimation of risk dierence
> library(Epi)
> library(xtable)
> dlt <- data.frame(rbind( c(29,14),c(20,22) ))
> colnames( dlt ) <- c("Better","Worse")
> dlt$trt <- c(1,0)
> b2<-glm(cbind(Better,Worse)~trt,
+ family=binomial(link="identity"),
+ data=dlt)
> xtable(round( ci.exp( b2,Exp=F ), digits=6 ))
Estimate 2.5% 97.5%
(Intercept) 0.48 0.33 0.63
trt 0.20 -0.01 0.40
Twenty percent more of the Diabetics with Dalterapin treatment are getting
better compared to Diabetics treated with placebo
19 / 20
Conclusion: What did we learn?
▶ Rates, their ratio and dierence can be analysed by Poisson regression
▶ In Poisson models the response can be either:
▶ case indicator d with offset = log(y), or
▶ case and person-years cbind(d,y) with poisreg-family (Epi-package)
▶ Both may be tted on either grouped data, or individual records.
▶ Binary outcome can be modeled with binary regression.

20 / 20