Labs

Statistical methods for population-based cancer survival analysis
Computing notes and exercises
Paul W. Dickman1 , Paul C. Lambert1,2 , Sandra Eloranta1 ,

Therese Andersson1 , Mark J Rutherford2 , Anna Johansson1 ,
Caroline E. Weibull1 , Sally Hinchliffe2 , Hannah Bower1 , Michael Crowther2
(1) Department of Medical Epidemiology and Biostatistics

Karolinska Institutet
Stockholm, Sweden
(2) Department of Health Sciences

Universtity of Leicester
Leicester, UK.
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
Cancer Research Centre, Cancer Institute of Iran,

Tehran University of Medical Sciences
1619 October 2016
1
2 CONTENTS
Contents
1 Notes on survival analysis using Stata 4
2 Downloading user-written Stata commands and data files 5

2.1 The quick and easy way . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Downloading the course files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Installing Stata user-written commands for relative survival . . . . . . . . . . . . 6
2.4 Relative survival in SAS and R . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3 Exercises 8
100. Hand calculation: Life table and Kaplan-Meier estimates of survival . . . . . . . . 8
101. Using Stata to validate the hand calculations done in question 100 . . . . . . . . 10
102. Comparing actuarial and Kaplan-Meier approaches with discrete-time data . . . . 12
103. Comparing cause-specific and all-cause survival . . . . . . . . . . . . . . . . . . . 13
104. Comparing estimates of cause-specific survival between periods; log rank test . . 15
110. Reviewing the Poisson regression example from the lecture notes (diet data) . . . 17
111. Model cause-specific mortality using Poisson regression . . . . . . . . . . . . . . . 19
112. Poisson regression with the diet data; choice of timescale . . . . . . . . . . . . . . 22
120. Model cause-specific mortality using Cox regression . . . . . . . . . . . . . . . . . 24
121. Examining the proportional hazards hypothesis (localised melanoma) . . . . . . . 26
122. Cox regression for all-cause mortality . . . . . . . . . . . . . . . . . . . . . . . . . 28
123. Examining the effect of sex on melanoma survival . . . . . . . . . . . . . . . . . . 29
124. Modelling the diet data using Cox regression . . . . . . . . . . . . . . . . . . . . . 30
125. Time-varying exposures the bereavement data . . . . . . . . . . . . . . . . . . . 31
130. Understanding splines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
131. Model cause-specific mortality using flexible parametric models . . . . . . . . . . 38
132. Flexible parametric models with time-dependent effects . . . . . . . . . . . . . . . 42
133. Modelling on other scales using stpm2 . . . . . . . . . . . . . . . . . . . . . . . . 46
140. Probability of death in a competing risks framework (cause-specific survival) . . . 48
150. Adjusted/standardized survival curves . . . . . . . . . . . . . . . . . . . . . . . . 55
180. Outcome-selective sampling designs . . . . . . . . . . . . . . . . . . . . . . . . . . 58
181. Calculating SMRs/SIRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
182. Calculating SMRs/SIRs using strs . . . . . . . . . . . . . . . . . . . . . . . . . . 66
200. Hand calculation of expected survival . . . . . . . . . . . . . . . . . . . . . . . . . 68
201. Life table estimates of relative survival using strs . . . . . . . . . . . . . . . . . . 69
202. Life table estimates of cause-specific survival using strs . . . . . . . . . . . . . . 71
CONTENTS 3
203. Period estimation of relative survival . . . . . . . . . . . . . . . . . . . . . . . . . 72

204. Evaluating period predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
210. Model excess mortality using Poisson regression . . . . . . . . . . . . . . . . . . . 74
211. Model excess mortality using Poisson regression with a smooth baseline . . . . . . 75
230. Model excess mortality using flexible parametric models . . . . . . . . . . . . . . 78
231. Non-linear effects in relative survival I Proportional hazards . . . . . . . . . . . 81
232. Non-linear effects in relative survival II Time-dependent effects . . . . . . . . . 83
240. Age-standardised estimates of relative survival (internal standard) . . . . . . . . . 85
241. Age-standardised estimates of relative survival . . . . . . . . . . . . . . . . . . . . 87
242. Age standardization using flexible parametric models . . . . . . . . . . . . . . . . 88
243. Age-standardised estimates of relative survival (external standard) . . . . . . . . 90
250. Probability of death in a competing risks framework (life table relative survival) . 91
251. Probability of death in a competing risks framework (relative survival model) . . 92
260. Fitting cure models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
261. Fitting cure models using flexible parametric models . . . . . . . . . . . . . . . . 95
280. Creating a popmort file from the Human Mortality Database . . . . . . . . . . . 97
281. Constructing a popmort file by modelling cohort data . . . . . . . . . . . . . . . . 99
282. Excess and avoidable deaths from life tables . . . . . . . . . . . . . . . . . . . . 101
283. Simulating relative survival . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
284. Estimating loss in expectation of life . . . . . . . . . . . . . . . . . . . . . . . . . 106
285. Missing covariate data (using official Stata commands) . . . . . . . . . . . . . . . 109
4 References 113
4
1 Notes on survival analysis using Stata
A general introduction to Stata (stataintro.pdf) can be downloaded from:

http://biostat3.net/download/.
If you are not familiar with Stata you should start by downloading and reading this introduction.
The same document includes an extensive description of the stset command that is central to
survival analysis.
In order to analyse survival data it is necessary to specify (at a minimum) a variable representing
the time at risk (e.g., survival time) and a variable specifying whether or not the event of interest
was observed (called the failure variable). Instead of specifying a variable representing time at
risk we may instead specify the entry and exit dates.
In many statistical software programs (such as SAS), these variables must be specified every
time a new analysis is performed. In Stata, these variables are specified once using the stset
command and then used for all subsequent survival analysis (st) commands (until the next
stset command). For example
. use melanoma
. stset surv_mm, failure(status==1)
The above code shows how we would stset the skin melanoma data in order to analyse cause-
specific survival with survival time in completed months (surv_mm) as the time variable. The
variable status takes the values 0=alive, 1=dead due to cancer, and 2=dead due to other
causes. We have specified that only status=1 indicates an event (death due to melanoma) so
Stata will consider observations with other values of status as being censored. If we wanted
to analyse observed survival (where all deaths are considered to be events) we could use the
following command
. stset surv_mm, failure(status==1,2)
Some of the Stata survival analysis (st) commands relevant to this course are given below.
Further details can be found in the manuals or online help.
stset Declare data to be survival-time data

stsplit Split time-span records
sts Generate, graph, list, and test the survivor and cumulative
hazard functions
strate Calculate person-time at risk and failure rates
stcox Estimate Cox proportional hazards model
streg Estimate parametric survival models
strs Life table estimation of relative survival
Once the data have been stset we can use any of these commands without having to specify
the survival time or failure time variables. For example, to plot the estimated cause-specific
survivor function by sex and then fit a Cox proportional hazards model with sex and calendar
period as covariates
. sts graph, by(sex)

. stcox sex year8594
5
2 Downloading user-written Stata commands and data files
Stata will be used throughout the course. This section describes how to download and install the
files required for the computing exercises (e.g., data files) as well as how to install user-written
commands for extending Stata. Standard Stata does not contain commands for relative survival,
so we must extend Stata with user-written commands. Note that there are two separate steps;
downloading the course files and installing the user-written commands. We have written an
automated script that does both these steps (see section 2.1) or if you prefer more control, there
are instructions in sections 2.2 and 2.3.
2.1 The quick and easy way
Enter the following at the Stata command line to download the course files (e.g., data files and
solution do files) and install all Stata user-written commands:
do http://www.pauldickman.com/survival/install_packages.do
NOTE: You do not need to change the working directory before running this command. This
will create a directory, c:\survival, and install the course files into that directory. If the
directory c:\survival already exists, or you prefer to install the files into another directory
then you will need to download install_packages.do, edit the directory reference, and then
run the file from within Stata.
2.2 Downloading the course files
You do not have to do this if you already used the quick and easy way described in section 2.1;
the course files will already be downloaded for you.
The course files (e.g., data files and solution do files) are distributed as a Stata package so should
be downloaded from within Stata. It is suggested that you create a new directory, change the
Stata working directory to the new directory (e.g., cd c:\survival\), and then download the
files. You can create a new directory in Windows Explorer or you can do it from within Stata
as follows.
mkdir c:\survival
cd c:\survival
Use the pwd command to confirm you are in the working directory you wish to use for the course
and then issue the following command from the Stata command line to install the course files.
net install http://www.pauldickman.com/survival/course_files, all replace
net install downloads the files and copies them to appropriate directories according to the
way Stata is setup. Ancillary files (e.g., PDF, XLS, DTA) are copied to the current working
directory; ADO and HLP files are installed into the appropriate directory according to the way
Stata is configured.
6
2.3 Installing Stata user-written commands for relative survival
You do not have to do this if you already used the quick and easy way described in section 2.1.
Standard Stata does not contain any commands for estimating and modelling relative survival
so we must extend Stata using commands written by users. Download and installation is done
within Stata. It is recommended that you change the Stata working directory to the course
directory (e.g., cd c:\survival\) before issuing these commands.
2.3.1 How can I check if these commands are already installed?
You can use the which command to check if (and where) a Stata command is installed.
. which stpm2
c:\ado\plus\s\stpm2.ado
*! version 1.6.3 14Jan2016
Use the adoupdate command to update previously installed user-written commands (note that
this is distinct from the update command that updates official Stata commands). Simply type
adoupdate, update to update all user-written commands.
2.3.2 strs - estimating and modelling relative survival
The strs command, written by Paul Dickman and Enzo Coviello can be downloaded by typing
the following:
. net install http://www.pauldickman.com/rsmodel/stata_colon/strs, all replace
Note that some of the data files are contained in both the strs and the course_files packages,
hence the need for the replace option. See http://pauldickman.com/rsmodel/stata_colon/
for further details about the command or read the Stata help file after installation. The command
is described in a Stata Journal article [1].
2.3.3 stpm2 - flexible parametric models
The stpm2 command, written by Paul Lambert and Patrick Royston, fits flexible parametric
survival models (so called Royston-Parmar models). Relative survival models can be fitted using
the bhazard() option. It is installed from within Stata using the following commands:
ssc install stpm2

ssc install rcsgen
The command is described in a Stata Journal article [2]. rcsgen is a command for generating
basis vectors for restricted cubic splines and is required by stpm2. Flexible parametric cure
models (fitted using an option to stpm2) are described in another Stata Journal article [3].
2.4 Relative survival in SAS and R 7
2.3.4 strsmix and strsnmix - cure models
To install strsmix and strsnmix (commands for fitting cure models) first type findit lambert cure
then click on the Stata Journal link followed by click to install. These commands are described
in a Stata Journal article [4].
2.3.5 Estimating probability of death in a competing risks framework
The stcompet command estimates the cumulative incidence function (CIF) non-parametrically.
The stcompadj command estimates the CIF using a competing risks analogue of the Cox model.
The stpm2cm command estimates the crude probabilities of death (i.e., CIF) after fitting a rel-
ative survival model using stpm2. The stpm2cif command estimates the CIF through postes-
timation after fitting a cause-specific competing risks model using stpm2.
ssc install stcompet

ssc install stcompadj
ssc install stpm2cm
ssc install stpm2cif
The stpm2cif command is described in a Stata Journal article [5].
2.4 Relative survival in SAS and R
Paul Dickman has written SAS code for estimating and modelling relative survival, see http:
//pauldickman.com/rsmodel/sas_colon/. Hermann Brenner and Colleagues have also pub-
lished SAS code, see http://www.imbe.med.uni-erlangen.de/issan/SAS/period/period.
htm. Maja Pohar has written an R package for relative survival that is easily found with search
engines.
8 EXERCISES
3 Exercises
100. Hand calculation: Life table and Kaplan-Meier estimates of survival

Using hand calculation (i.e., using a spreadsheet program or pen, paper, and a calculator)
estimate the cause-specific survivor function for the sample of 35 patients diagnosed with
colon carcinoma (see the table below) using both the Kaplan-Meier method (up to at least
30 months) and the actuarial method (at least the first 5 annual intervals).
In the lectures we estimated the observed survivor function (i.e. all deaths were considered
to be events) using the Kaplan-Meier and actuarial methods; your task is to estimate the
cause-specific survivor function (only deaths due to colon carcinoma are considered events)
using the same data. The next page includes some hints to help you get started.
Age Clinical dx date Surv. time

ID Sex at dx stage mmyy mm yy Status
1 male 72 Localised 2.89 2 0 Dead - other
2 female 82 Distant 12.91 2 0 Dead - cancer
3 male 73 Distant 11.93 3 0 Dead - cancer
5 male 67 Localised 5.89 7 0 Dead - cancer
6 male 74 Regional 7.92 8 0 Dead - cancer
9 male 64 Localised 11.94 13 1 Alive
10 female 70 Localised 10.94 14 1 Alive
11 female 83 Localised 7.90 19 1 Dead - other
15 male 70 Regional 9.93 27 2 Alive
24 male 54 Regional 6.85 46 3 Dead - cancer
32 female 80 Localised 5.87 102 8 Dead - cancer
33 male 66 Localised 1.86 103 8 Dead - other
35 female 56 Distant 12.86 108 9 Alive
9
ACTUARIAL APPROACH
We suggest you start with the actuarial approach. Your task is to construct a life table
with the following structure.
time l d w l0 p S(t)
[0-1) 35
[1-2)
[2-3)
[3-4)
[4-5)
[5-6)
We have already entered l1 (number of people alive at the start of interval 1). The next
step is to add the number who experienced the event (d) and the number censored (w)
during the first year. From l, d, and w you will then be able to calculate l0 (effective num-
ber at risk), followed by p (conditional probability of surviving the interval) and finally
S(t), the cumulative probability of surviving from time zero until the end of the interval.
KAPLAN-MEIER APPROACH
To estimate survival using the Kaplan-Meier approach you will find it easiest to add a line
to the table at each and every time there is an event or censoring. We should use time in
months. The first time at which there is an event or censoring is time equal to 2 months.
The trick is what to do when there are both events and censorings at the same time.
time # at risk d w p S(t)
2 35
10 EXERCISES
101. Using Stata to validate the hand calculations done in question 100
We will now use Stata to reproduce the same analyses done by hand calculation in question
100 although you can do this part without having done the hand calculations, since this
question also serves as an introduction to survival analysis using Stata. Our aim is to
estimate the cause-specific survivor function for the sample of 35 patients diagnosed with
colon carcinoma using both the Kaplan-Meier method and the actuarial method. In the
lectures we estimated the all-cause survivor function (i.e. all deaths were considered to
be events) using the Kaplan-Meier and actuarial methods whereas we will now estimate
the cause-specific survivor function (only deaths due to colon carcinoma are considered
events).
After starting Stata, you will first have to specify the data set you wish to analyse, that is
. use colon_sample, clear
Stata will search for this file in the current working directory. The pwd command will return
the name of the current working directory. If you need to change to another directory you
can use, for example, cd c:\survival\. The describe command will return a summary
of the data set structure (e.g., variable names) whereas the list command will display
the values of variables.
In order to use the Stata ltable command (life table estimates of the survivor function)
we must construct a new variable indicating whether the observation period ended with an
event (the new variable is assigned code 1) or censoring (the new variable is assigned code
0). We will call this new variable csr_fail (cause-specific failure). The ltable command
is not a standard Stata survival analysis (st) command and does not require that the data
be stset.
. recode status (1=1) (nonmissing=0), gen(csr_fail)
There are many ways to create the new variable, the above approach is preferred because
missing values of status will remain missing. Even though we dont have any missing
values, it is good programming practice to always write code that will handle missing
values appropriately.
The following command will give the actuarial estimates
. ltable surv_yy csr_fail
Alternatively, we could use
. ltable surv_mm csr_fail, interval(12)
Before most Stata survival analysis commands can be used (ltable is an exception) we
must first stset the data using the stset command (see Section 1).

11
A listing of the Kaplan-Meier estimates is then obtained as follows
. sts list
To graph the Kaplan-Meier estimates
. sts graph
Note that we only have to stset the data once. You can also tell Stata to show the number
at risk either on the curve or in a table.
. sts graph, atrisk

. sts graph, risktable
Titles and axis labels can also be specified.
. sts graph, risktable ///

title(Kaplan-Meier estimates of cause-specific survival) ///
xtitle(Time since diagnosis in months)
12 EXERCISES
102. Localised melanoma: Comparing actuarial and Kaplan-Meier approaches with

discrete time data
The aim of this exercise is to examine the effect of heavily grouped data (i.e., data with lots
of ties) on estimates of survival made using the Kaplan-Meier method and the actuarial
method.
For the patients diagnosed with localised skin melanoma, use Stata to estimate the 10-year
cause-specific survival proportion. Use both the Kaplan-Meier method and the actuarial
method. Do this both with survival time recorded in completed years and survival time
recorded in completed months. That is, you should obtain 4 separate estimates of the
10-year cause-specific survival proportion to complete the cells of the following table. The
purpose of this exercise is to illustrate small differences between the two methods when
there are large numbers of ties.
In order to reproduce the results in the printed solutions youll need to restrict to localised
stage (stage==1) and estimate cause-specific survival (status==1 indicates an event).
Look at the Stata code in the previous questions if you are unsure.
Actuarial Kaplan-Meier
Years
Months
(a) Of the two estimates (Kaplan-Meier and actuarial) made using time recorded in years,
which do you think is the most appropriate and why?
[HINT: Consider how each of the methods handle ties.]
(b) Which of the two estimates (Kaplan-Meier or actuarial) changes most when using
survival time in months rather than years? Why?
13
103. Melanoma: Comparing survival proportions and mortality rates by stage for
cause-specific and all-cause survival
The purpose of this exercise is to study survival of the patients using two alternative
measures - survival proportions and mortality rates. A second purpose is to study the
difference between cause-specific and all-cause survival.
. use melanoma, clear

(a) Plot estimates of the survivor function and hazard function by stage.
. sts graph, by(stage)
. sts graph, hazard by(stage)
By default, the sts graph command plots Kaplan-Meier estimates of survival. If we
add the hazard option it shows estimates of the hazard function. Does it appear that
stage is associated with patient survival?
Stata tip: You may have found that each time you produce a graph Stata overwrites
the previous graph in the graph window. You can instruct Stata to open each graph
in a separate window by naming the graphs. This will give you the possibility to
compare graphs side by side.
. sts graph, by(stage) name(survival)
. sts graph, by(stage) name(hazard) hazard
You can use set autotabgraphs to control whether multiple graphs are created as
tabs within one window or as separate windows. Issue the following command to
make Stata present graphs as tabs within a single window (and store the setting
permanently).
set autotabgraphs on, permanently
(b) Estimate the mortality rates for each stage using, for example, the strate command.
. strate stage
What are the units of the estimated rates?
[The strate command, as the name suggests, is used to estimates rates. Look at the
help pages if you are not familiar with the command.]
(c) If you havent already done so, estimate the mortality rates for each stage per 1000
person-years of follow-up.
[HINT: consider the scale() option to stset and the per() option to strate.]
(d) Study whether survival is different for males and females (both by plotting the sur-
vivor function and by tabulating mortality rates).
. sts graph, hazard by(sex)
Is there a difference in survival between males and females? If yes, is the difference
present throughout the follow up?
14 EXERCISES
(e) The plots you made above were based on cause-specific survival (i.e., only deaths due
to cancer are counted as events, deaths due to other causes are censored). In the next
part of this question we will estimate all-cause survival (i.e., any death is counted as
an event). First, however, study the coding of vital status and tabulate vital status
by age group.
How many patients die of each cause? Does the distribution of cause of death depend
on age?
. codebook status
. tab status agegrp
(f) To get all-cause survival, specify all deaths (both cancer and other) as events in the
stset command.
Now plot the survivor proportion for all-cause survival by stage. We name the graph
to be able to separate them in the graph window. Is the survivor proportion different
compared to the cause-specific survival you estimated above? Why?
. sts graph, by(stage) name(anydeath, replace)
(g) It is more common to die from a cause other than cancer in older ages. How does
this impact the survivor proportion for different stages? Compare cause-specific and
all-cause survival by plotting the survivor proportion by stage for the oldest age group
(75+ years) for both cause-specific and all-cause survival. We suggest you copy the
code from the PDF file into the Stata do editor and run the code from there.
. sts graph if agegrp==3, by(stage) ///
name(cancerdeath_75, replace) subtitle("Cancer")
. sts graph if agegrp==3, by(stage) ///
name(anydeath_75, replace) subtitle("All cause")
. graph combine cancerdeath_75 anydeath_75
(h) Now estimate both cancer-specific and all-cause survival for each age group.
. sts graph, by(agegrp) name(anydeathbyage, replace) subtitle("All cause")

. sts graph, by(agegrp) name(cancerdeathbyage, replace) subtitle("Cancer")
. graph combine anydeathbyage cancerdeathbyage

Are there bigger differences between the age groups for cause-specific or for all-cause
survival?
15
104. Localised melanoma: Comparing estimates of cause-specific survival between

periods; first graphically and then using the log rank test
We will now analyse the full data set of patients diagnosed with localised skin melanoma.
Use Stata to estimate the cause-specific survivor function, using the Kaplan-Meier method
with survival time in months, separately for each of the two calendar periods 19751984
and 19851994. The following commands can be used
. use melanoma if stage == 1, clear

. sts graph, by(year8594)
The variable year8594 takes the value 1 for patients diagnosed 19851994 and 0 for those
diagnosed 19751984.
(a) Without making reference to any formal statistical tests, does it appear that patient
survival is superior during the most recent period?
(b) The following commands can be used to plot the hazard function (instantaneous
mortality rate):
. sts graph, hazard by(year8594)
i. At what point in the follow-up is mortality highest?
ii. Does this pattern seem reasonable from a clinicial/biological perspective? [HINT:
Consider the disease with which these patients were classified as being diagnosed
along with the expected fatality of the disease as a function of time since diag-
nosis.]
(c) Use the log rank test to determine whether there is a statistically significant difference
in patient survival between the two periods. The following command can be used:
. sts test year8594
What do you conclude?
An alternative test is the generalised Wilcoxon, which can be obtained as follows
. sts test year8594, wilcoxon
Havent heard of the log rank (or Wilcoxon) test? Its possible you may reach this
exercise before we cover the details of these tests during lectures. You should nev-
ertheless do the exercise and try and interpret the results. Both of these tests (the
log rank and the generalised Wilcoxon) are used to test for differences between the
survivor functions. The null hypothesis is that the survivor functions are equivalent
for the two calendar periods (i.e., patient survival does not depend on calendar period
of diagnosis).
16 EXERCISES
(d) Estimate cause-specific mortality rates for each age group, and graph Kaplan-Meier
estimates of the cause-specific survivor function for each age group. Are there differ-
ences between the age groups? Is the interpretation consistent between the mortality
rates and the survival proportions?
. strate agegrp, per(1000)
. sts graph, by(agegrp)
What are the units of the estimated hazard rates? HINT: look at how you defined
time when you stset the data.
(e) Repeat some of the previous analyses after using the scale() option to stset to rescale
time from months to years. This is equivalent to dividing the time variable by 12
so all analyses will be the same except the units of time will be different (e.g., the
graphs will have different labels).
. stset surv_mm, failure(status==1) scale(12)
. sts graph, by(agegrp)
. strate agegrp, per(1000)
(f) Study whether there is evidence of a difference in patient survival between males and
females. Estimate both the hazard and survival function and use the log rank test to
test for a difference.
17
110. Diet data: tabulating incidence rates and modelling with Poisson regression
Load the diet data and stset the data using time-on-study as the timescale.
. use diet, clear

. stset dox, id(id) fail(chd) origin(doe) scale(365.24)
(a) Use the strate command to tabulate CHD incidence rates per 1000 person-years for
each category of hieng. Calculate (by hand) the ratio of the two incidence rates.
(b) Use the command poisson to find the incidence rate ratio for the high energy group
compared to the low energy group and compare the estimate to the one you obtained
in the previous question:
. poisson chd hieng, e(y) irr
NOTE: Rates are calculated as events/person-time so when modelling rates we need

to let Stata know both of these quantities. chd is the event indicator and y is the
person time at risk for each individual. The irr option results in the estimates being
presented as estimated incidence rate ratios rather than parameter estimates (log
incidence rate ratios).
(c) Grouping the values of total energy into just two groups does not tell us much about
how the CHD rate changes with total energy. It is a useful exploratory device, but
to look more closely we need to group the total energy into perhaps 3 or 4 groups.
In this example we shall use the cut points 1500, 2500, 3000, 4500. To check if these
cutpoints seem reasonable, type:
. histogram energy, normal
. sum energy, detail
(d) Use the commands
. egen eng3=cut(energy), at(1500, 2500, 3000, 4500)
. tabulate eng3
to create a new variable eng3 coded 1500 for values of energy in the range 15002499,
2500 for values in the range 25002999, and 3000 for values in the range 30004500.
(e) To estimate and plot the rates for different levels of eng3 try
. strate eng3, per(1000) graph
Calculate (by hand) the ratio of rates in the second and third levels to the first level.
(f) Create your own indicator variables for the three levels of eng3 with
. tabulate eng3, gen(X)
(g) Check the indicator variables with
. list energy eng3 X1 X2 X3 if eng3==1500
18 EXERCISES
(h) Use poisson to compare the second and third levels with the first, as follows:
. poisson chd X2 X3, e(y) irr
Compare your estimates with those you obtained in 110e.

(i) Use poisson to compare the first and third levels with the second.
(j) Repeat the analysis comparing the second and third levels with the first but this time
have Stata create the indicators automatically via the i. syntax. That is
. poisson chd i.eng3, e(y) irr
(k) Without using st commands, calculate the total number of events during follow-
up, person-time at risk, and the crude incidence rate (per 1000 person-years), for
example with Stata commands for descriptive statistics (e.g., summarize). Confirm
your answer using strate or stptime.
[HINT: Remember that the total number of person-years is the number of persons at
risk multiplied with the mean follow up time among those persons.]
19
111. Localised melanoma: model cause-specific mortality with Poisson regression
In this exercise we model, using Poisson regression, cause-specific mortality of patients

diagnosed with localised (stage==1) melanoma.
In exercise 120 we model cause-specific mortality using Cox regression and in exercise 131
we use flexible parametric models. The aim is to illustrate that these three methods are
very similar. Cox regression cannot be used to model excess mortality, so we use Poisson
regression (exercise 210) or flexible parametric models (exercise 230).
The aim of these exercises is to explore the similarities and differences to these three
approaches to modelling. We will be comparing the results (and their interpretation) as
we proceed through the exercises so you may wish to save your commands in a do file to
facilitate comparison.
The following commands can be used to load and stset the data.
. use melanoma if stage==1, clear

. stset surv_mm, failure(status==1) scale(12) id(id)
(a) Plot Kaplan-Meier estimates of cause-specific survival as a function of calendar period

of diagnosis.
. sts graph, by(year8594)
i. During which calendar period (the early or the latter) is survival best?
ii. Now plot the estimated hazard function (cause-specific mortality rate) as a func-
tion of calendar period of diagnosis.
. sts graph, by(year8594) hazard
During which calendar period (the early or the latter) is mortality the lowest?
iii. Is the interpretation (with respect to how prognosis depends on period) based on
the hazard consistent with the interpretation of the survival plot?
(b) Use the strate command to estimate the cause-specific mortality rate for each cal-
endar period.
. strate year8594, per(1000)
During which calendar period (the early or the latter) is mortality the lowest? Is this
consistent with what you found earlier? If not, why the inconsistency?
(c) The reason for the inconsistency between parts 111a and 111b was confounding by
time since diagnosis. The comparison in part 111a was adjusted for time since di-
agnosis (since we compare the differences between the curves at each point in time)
whereas the comparison in part 111b was not. Understanding this concept is central
to the remainder of the exercise so please ask for help if you dont follow.
Two approaches for controlling for confounding are restriction and statistical ad-
justment. We will first use restriction to control for confounding. That is we will
stset the data again but use the exit(time 120) option to restrict the potential
follow-up time to a maximum of 120 months. Individuals who survive more than 120
months are censored at 120 months
. stset surv_mm, failure(status==1) scale(12) id(id) exit(time 120)
20 EXERCISES
i. Use the strate command to estimate the cause-specific mortality rate for each
calendar period.
. strate year8594, per(1000)
During which calendar period (the early of the latter) is mortality the lowest? Is
this consistent with what you found in part 111b?
ii. Calculate by hand the ratio (8594/7584) of the two mortality rates (i.e., a
mortality rate ratio) and interpret the estimate (i.e., during which period is
mortality higher/lower and by how much).
iii. Now use Poisson regression to estimate the same mortality rate ratio.
. streg year8594, dist(exp)
NOTE: streg is one of several Stata commands for performing Poisson regression.
The model could also be fitted using the poisson or glm commands.
. gen risktime=_t-_t0
. poisson _d year8594 if _st==1, exp(risktime) irr
. glm _d year8594 if _st==1, family(poisson) eform lnoffset(risktime)
However, if you have stset/stsplit the data it is recommended that you use streg
since streg understands and respects the internal st variables (_st, _t, _t0, and
_d). In particular, trimmed person-time will be ignored by streg but not by the
poisson command.
Strictly speaking, streg fits parametric survival models. A parametric survival model
assuming survival times are exponentially distributed (dist(exp)) implies a constant
hazard and a Poisson process for the number of events (i.e., Poisson regression).
(d) In order to adjust for time since diagnosis (i.e., adjust for the fact that we expect mor-
tality to depend on time since diagnosis) we need to split the data by this timescale.
We will restrict our analysis to mortality up to 10 years following diagnosis.
. stsplit fu, at(0(1)10) trim
NOTE: The trim option instructs Stata to ignore time-at-risk outside the interval
[0,10] (i.e., after 10 years subsequent to diagnosis). Since we have already made this
restriction using stset there should not be any time trimmed.
(e) Now tabulate (and produce a graph of) the rates by follow-up time.
. strate fu, per(1000) graph
Mortality appears to be quite low during the first year of follow-up. Does this seem
reasonable considering the disease with which these patients have been diagnosed?
(f) Compare the plot of the estimated rates to a plot of the hazard rate as a function of
continuous time.
. sts graph, hazard
Is the interpretation similar? Do you think it is sufficient to classify follow-up time

into annual intervals or might it be preferable to use, for example, narrower intervals?
21
(g) Use Poisson regression to estimate incidence rate ratios as a function of follow-up
time.
. streg i.fu, dist(exp)
Does the pattern of estimated incident rate ratios mirror the pattern you observed in
the plots?
(h) Now estimate the effect of calendar period of diagnosis while adjusting for time since
diagnosis. Before fitting this model, predict what you expect the estimated effect to
be (i.e., will it be higher, lower, or similar to the value of 0.8831852 we obtained in
part 111c.
. streg i.fu year8594, dist(exp)
Is the estimated effect of calendar period of diagnosis consistent with what you ex-
pected? Add an interaction between follow-up and calendar period of diagnosis and
interpret the results.
(i) Now control for age, sex, and calendar period.
. streg i.fu i.agegrp year8594 sex, dist(exp)
i. Interpret the estimated hazard ratio for the parameter labelled agegrp 2, in-
cluding a comment on statistical significance.
ii. Is the effect of calendar period strongly confounded by age and sex? That is,
does the inclusion of sex and age in the model change the estimate for the effect
of calendar period?
iii. Perform a Wald test of the overall effect of age and interpret the results.
. test 1.agegrp 2.agegrp 3.agegrp
(j) Is the effect of sex modified by calendar period (whilst adjusting for age and follow-
up)? Fit an appropriate interaction term to test this hypothesis.
(k) Based on the interaction model you fitted in exercise 111j, estimate the hazard ratio
for the effect of sex (with 95% confidence interval) for each calendar period.
ADVANCED: Do this with each of the following methods and confirm that the results
are the same:
i. Using hand-calculation on the estimates from exercise 111j.
ii. Using the estimates from exercise 111j and the lincom command.
. lincom 2.sex + 1.year8594#2.sex, eform
iii. Creating appropriate dummy variables that represent the effects of sex for each
calendar period.
. gen sex_early=(sex==2)*(year8594==0)
. gen sex_latter=(sex==2)*(year8594==1)
. streg i.fu i.agegrp year8594 sex_early sex_latter, dist(exp)
iv. Using Stata 11 syntax to repeat the previous model.
. streg i.fu i.agegrp i.year8594 year8594#sex, dist(exp)
(l) Now fit a separate model for each calendar period in order to estimate the hazard
ratio for the effect of sex (with 95% confidence interval) for each calendar period.
Why do the estimates differ from those you obtained in the previous part?
. streg i.fu i.agegrp sex if year8594==0, dist(exp)
. streg i.fu i.agegrp sex if year8594==1, dist(exp)
Can you fit a single model that reproduces the estimates you obtained from the
stratified models? Try:
. streg i.fu##year8594 i.agegrp##year8594 year8594##sex, dist(exp)
22 EXERCISES
112. Diet data: Using Poisson regression to study the effect of energy intake ad-
justing for confounders on two different timescales
Use Poisson regression to study the association between energy intake (hieng) and CHD
adjusted for potential confounders (job, BMI). We know that people who expend a lot of
energy (i.e., are physically active) require a higher energy intake. We do not have data on
physical activity but we are hoping that occupation (job) will serve as a surrogate measure
of work-time physical activity (conductors on London double-decker busses expend energy
walking up and down the stairs all day).
Fit models both without adjusting for time and by adjusting for attained age (you will
need to split the data) and time-since-entry and compare the results.
(a) Rates can be modelled on different timescales, e.g., attained age, time-since-entry,
calendar time. Plot the CHD incidence rates both by attained age and by time-since-
entry. Is there a difference? Do the same for CHD hazard by different energy intakes
(hieng).
. use diet, clear
.* Timescale: Attained age

. stset dox, id(id) fail(chd) origin(dob) enter(doe) scale(365.24)
. sts graph, hazard
. sts graph, by(hieng) hazard
.* Timescale: Time-since-entry
. stset dox, id(id) fail(chd) origin(doe) enter(doe) scale(365.24)
. sts graph, hazard
. sts graph, by(hieng) hazard
(b) Model the rate using Poisson regression, without adjusting for any timescale. What
is the effect of hieng on CHD? What assumption does this model make on the shape
of the underlying incidence rate over time?
(c) Adjust for BMI and job. Is there evidence that the effect of energy intake on CHD
is confounded by BMI and job?
. gen bmi=weight/(height/100*height/100)
. poisson chd hieng job bmi, e(y) irr
(d) Firstly, lets adjust for the timescale attained age. To do this in Poisson regression
you must split the data on timescale age. First use stset (with origin date of birth)
and then use stsplit to generate agebands.
. stset dox, id(id) fail(chd) origin(dob) enter(doe) scale(365.24)

. stsplit ageband, at(30,50,60,72) trim
. list id _t0 _t ageband y in 1/10
As the poisson command is not an st command, you must keep track of the risktime
yourself. Why is the y variable not correct anymore? Generate a new variable,
risktime, which contains the risktime for each split record.
23
. list id _t0 _t ageband y risktime in 1/10
You must also keep track of the event variable, as chd will not be valid after the split.
. tab ageband chd, missing
. tab ageband _d, missing
Now fit the model for CHD, both without and with the adjustment for job and bmi.
Is the effect of hieng on CHD confounded by age, BMI or job?
. poisson _d hieng i.ageband, e(risktime) irr
. poisson _d hieng i.job bmi i.ageband, e(risktime) irr
What assumption is being made about the shape of the baseline hazard (HINT: the
baseline hazard takes the shape of the timescale)?
(e) Secondly, do the same analysis, but now adjust for the timescale time-since-entry.
(You must read the data in again, as you now want to split on another timescale.
This is strictly not necessary, but to avoid mistakes it is generally a good idea to start
over again.)
. use diet, clear

. gen bmi=weight/(height/100*height/100)
Specify time-since-entry as the timescale by specifying date of entry as the time origin.
. stset dox, id(id) fail(chd) origin(doe) enter(doe) scale(365.24)
. stsplit fuband, at(0,5,10,15,22) trim

. list id _t0 _t fuband y in 1/10
. list id _t0 _t fuband y risktime in 1/10
. tab fuband chd, missing

. tab fuband _d, missing
. poisson _d hieng i.fuband, e(risktime) irr

. poisson _d hieng i.job bmi i.fuband, e(risktime) irr
Compare the results with the analysis adjusted for attained age. Are there any
differences? Why (or why not)? Go back to the graphs at the beginning of the
exercise and look for explanations.
(f) Repeat the exercise using streg. What is the advantage/disadvantage of using streg?
24 EXERCISES
120. Localised melanoma: modelling cause-specific mortality using Cox regression
In exercise 111 we modelled the cause-specific mortality of patients diagnosed with localised
melanoma using Poisson regression. We will now model cause-specific mortality using Cox
regression and compare the results to those we obtained using the Poisson regression model.
To fit a Cox proportional hazards model (for cause-specific survival) with calendar period
as the only explanatory variable, the following commands can be used. Note that we are
censoring all survival times at 120 months (10 years) in order to facilitate comparisons
with the Poisson regression model in exercise 111.
. use melanoma
. keep if stage == 1
. stset surv_mm, failure(status==1) exit(time 120)
. stcox year8594
(a) Interpret the estimated hazard ratio, including a comment on statistical significance.
(b) (This part is more theoretical and is not required in order to understand the remaining
parts.)
Stata reports a Wald test of the null hypothesis that survival is independent of cal-
endar period. The test statistic (and associated P-value) is reported in the table of
parameter estimates (labelled z). Under the null hypothesis, the test statistic has a
standard normal (Z) distribution, so the square of the test statistic will have a chi
square distribution with one degree of freedom.
Stata also reports a likelihood ratio test statistic of the null hypothesis that none of
the parameters in the model are associated with survival (labelled LR chi2(1)). In
general, this test statistic will have a chi-square distribution with degrees of freedom
equal to the number of parameters in the model. For the current model, with only one
parameter, the test statistic has a chi square distribution with one degree of freedom.
Compare these two test statistics with each other and with the log rank test statistic
(which also has a 21 distribution) calculated in question 104c (you should, however,
recalculate the log rank test since we have restricted follow-up to the first 10 years in
this exercise). Would you expect these test statistics to be similar? Consider the null
and alternative hypotheses of each test and the assumptions involved with each test.
(c) Now include sex and age (in categories) in the model.
. stcox sex year8594 i.agegrp
i. Interpret the estimated hazard ratio for the parameter labelled agegrp 2, in-
cluding a comment on statistical significance.
ii. Is the effect of calendar period strongly confounded by age and sex? That is,
does the inclusion of sex and age in the model change the estimate for the effect
of calendar period?
iii. Perform a Wald test of the overall effect of age and interpret the results.
. test 1.agegrp 2.agegrp 3.agegrp
25
(d) Perform a likelihood ratio test of the overall effect of age and interpret the results.
The following commands can be used
. est store A
. stcox sex year8594
. lrtest A
Compare your findings to those obtained using the Wald test. Are the findings
similar? Would you expect them to be similar?
(e) The model estimated in question 120c is similar to the model estimated in ques-
tion 111i.
i. Both models adjust for sex, year8594, and i.agegrp but the Poisson regres-
sion model in question 111i appears to adjust for an additional variable (i.fu).
Is the Poisson regression model adjusting for an additional factor? Explain.
ii. Would you expect the parameter estimate for sex, period, and age to be similar
for the two models? Are they similar?
iii. Do both models assume proportional hazards? Explain.
(f) Following is some code for estimating and comparing the Cox and Poisson regression
models.
use melanoma if stage==1, clear
stset surv_mm, failure(status==1) id(id) exit(time 120)
stcox year8594 sex i.agegrp
est store Cox
/* split on time since diagnosis (1-year intervals) */

stsplit fu, at(0(12)120) trim
streg i.fu year8594 sex i.agegrp, dist(exp)

est store Poisson
est table Cox Poisson, eform equations(1)
(g) ADVANCED: By splitting at each failure time we can estimate a Poisson regression
model that is identical to the Cox model. Code is available in the file q120.do on
the course website. This model might take several minutes to estimate and you may
need to reset the values of memory and matsize.
(h) ADVANCED: Split the data finely (e.g., 3-month intervals) and model the effect of
time using a restricted cubic spline.
use melanoma if stage==1, clear
stset surv_mm, failure(status==1) id(id) exit(time 120)
/* split on time since diagnosis (1-month intervals) */
stsplit fu, at(0(1)120) trim
/* Create basis for restricted cubic spline */
mkspline fu_rcs=fu, cubic
streg fu_rcs* year8594 sex i.agegrp, dist(exp)
predict xb, xb
twoway line xb fu if year8594==0 & sex==1 & agegrp==1, sort
26 EXERCISES
121. Examining the proportional hazards hypothesis (localised melanoma)
(a) For the localised melanoma data with 10 years follow-up, plot the instantaneous
cause-specific hazard for each calendar period. The following commands can be used
. use melanoma if stage == 1, clear
. stset surv_mm, failure(status==1) id(id) exit(time 120) scale(12)
. sts graph, hazard by(year8594)
Make a rough estimate of the hazard ratio for patients diagnosed 198594 to those
diagnosed 197584. In part (d) you will fit a Cox model and check your estimate.
(b) Now plot the instantaneous cause-specific hazard for each calendar period using a log
scale for the y axis (use the option yscale(log)). What would you expect to see if
a proportional hazards assumption were appropriate? Do you see it?
(c) Another graphical way of checking the proportional hazards assumption is to plot
the log cumulative cause specific hazard function for each calendar period. These
plots were not given extensive coverage in the lectures, so attempt this if you like or
continue to part (d). The command for plotting this function is
. stphplot, by(year8594)
What would you expect to see if a proportional hazards assumption were appropriate?
Do you see it?
(d) Compare your estimated hazard ratio from part (a) with the one from a fitted Cox
model with calendar period as the only explanatory variable. Are they similar?
(e) Now fit a more complex model and use graphical methods to explore the assumption
of proportional hazards by calendar period. For example,
. stcox sex i.year8594 i.agegrp
. estat phtest, plot(1.year8594)

(f) Do part (a)(e) but now for the variable agegrp. What are your conclusions regarding
the assumption of proportional hazards?
(g) Now formally test the assumption of proportional hazards using
. stcox sex i.year8594 i.agegrp
. estat phtest, detail
Are your conclusions from the test coherent with your conclusions from the graphical
assessments?
(h) Estimate separate age effects for the first two years of follow-up (and separate esti-
mates for the remainder of the follow-up) while controlling for sex and period. Do
the estimates for the effect of age differ between the two periods of follow-up?
There are two ways to fit time-varying effects: 1) the tvc option in stcox or 2) by
splitting on time using stsplit.
Using tvc:
. tab(agegrp), gen(agegrp)
. stcox sex year8594 agegrp2 agegrp3 agegrp4, ///
tvc(agegrp2 agegrp3 agegrp4) texp(_t>=2)
27
Using stsplit:
. stsplit fuband, at(0,2)
. list id _t0 _t fu in 1/10
. stcox sex year8594 i.agegrp##i.fuband
These are simply two alternative syntaxes for fitting the same model with the same
parameterizations. They give the so-called default parameterizations for interaction
effects. We see effects of age (i.e., the hazard ratios) for the period 02 years sub-
sequent to diagnosis along with the interaction effects. An advantage of the default
parameterisation is that one can easily test the statistical significance of the inter-
action effects. Before going further, test whether the age*follow-up interaction is
statistically significant (using a Wald and/or LR test).
(i) Often we wish to see the effects of exposure (age) for each level of the modifier (time
since diagnosis). That is, we would like to complete the table below with relevant
hazard ratios. To get the effects of age for the period 2+ years after diagnosis, using
the default parametrization, we must multiply the hazard ratios for 02 years by
the appropriate interaction effect. Now lets reparameterise the model to directly
estimate the effects of age for each level of time since diagnosis. This is easily done
in Stata (version 11 or later) using single #s
. stcox sex year8594 i.fuband i.fuband#i.agegrp
02 years 2+ years
Agegrp1 1.00 1.00
Agegrp2
Agegrp3
Agegrp4
Fill in the table above. Does the effect of age appear different before and after 2
years?
(j) ADVANCED: Fit an analogous Poisson regression model. Are the parameter esti-
mates similar? HINT: You will need to split the data by time since diagnosis.
28 EXERCISES
122. Cox regression with observed (all-cause) mortality as the outcome
Now fit a model to the localised melanoma data where the outcome is observed survival
(i.e. all deaths are considered to be events).
. stset surv_mm, failure(status==1,2) exit(time 120)

. keep if stage==1
(a) Interpret the estimated hazard ratio for the parameter labelled 2.agegrp, including
a comment on statistical significance.
(b) On comparing the estimates between the observed and cause-specific survival models
it appears that only the parameters for age have changed substantially. Can you
explain why the estimates for the effect of age would be expected to change more
than the estimates of the effect of sex and period?
29
123. Cox model for cause-specific mortality for melanoma (all stages)
Use Cox regression to model the cause-specific survival of patients with skin melanoma
(including all stages).
(a) First fit the model with sex as the only explanatory variable. Does there appear to
be a difference in survival between males and females?
(b) Is the effect of sex confounded by other factors (e.g. age, stage, subsite, period)?
After controlling for potential confounders, does there still appear to be a difference
in survival between males and females?
(c) Consider the hypothesis that there exists a class of melanomas where female sex hor-
mones play a large role in the etiology. These hormone related cancers are diagnosed
primarily in women and are, on average, less aggressive (i.e., prognosis is good). If
such a hypothesis were true we might expect the effect of sex to be modified by age
at diagnosis (e.g., pre versus post menopausal). Test whether this is the case.
(d) Decide on a most appropriate model for these data. Be sure to evaluate the propor-
tional hazards assumption.
30 EXERCISES
124. Modelling the diet data using Cox regression
(a) Fit the following Poisson regression model to the diet data (we fitted this same model
in question 110).
. use diet, clear
Now fit the following Cox model.
. stset dox, id(id) fail(chd) entry(doe) origin(doe) scale(365.24)
. stcox hieng
i. On what scale are we measuring time ? That is, what is the timescale?
ii. Is it correct to say that both of these models estimate the effect of high energy
on CHD without controlling for any potential confounders? If not, how are these
models conceptually different?
iii. Would you expect the parameter estimates for these two models to be very dif-
ferent? Is there a large difference?
(b) stset the data with attained age as the timescale and refit the Cox model. Is the
estimate of the effect of high energy different? Would we expect it to be different?
31
125. Estimating the effect of a time-varying exposure the bereavement data
These data were used to study a possible effect of marital bereavement (loss of husband
or wife) on allcause mortality in the elderly (see Clayton & Hills [6], 32.2). The dataset
was extracted from a larger follow-up study of an elderly population and concerns subjects
whose husbands or wives were alive at entry to the study. Thus all subjects enter as not
bereaved but may become bereaved at some point during followup. The variable dosp
records the date of death of each subjects spouse and takes the value 1/1/2000 where this
has not yet happened.
(a) Load the data with
. use brv, clear
. desc
To see how the coding works for couples try
. list id sex doe dosp dox fail if couple==3
for a couple, both of whom die during followup. Draw a picture showing the follow
up for both subjects, and mark the dates of entry exit and death of spouse on it.
Try
for a couple, one of whom dies during followup,
for a couple, neither of whom die during followup, and
for a couple where only data on one individual is available.

(b) Set the st variables, calculate the mortality rate per 1000 years for men and for
women, and find the rate ratio comparing women (coded 2) with men (coded 1),
using
. stset dox, fail(fail) origin(dob) entry(doe) scale(365.24) id(id)
. strate sex, per(1000)
. streg sex, dist(exp)
i. What dimension of time did we use as the timescale when we stset the data?
Do you think this is a sensible choice?
ii. Which gender has the highest mortality? Is this expected?
iii. Could age be a potential confounder? Does age at entry differ between males
and females? Later we will estimate the rate ratio while controlling for age.
(c) Breaking records into pre and post bereavement. In these data a subject
changes exposure status from not bereaved to bereaved when his or her spouse dies.
The first stage of the analysis therefore is to partition each followup into a record de-
scribing the period of follow-up prebereavement and (for subjects who were bereaved
during the study) the period postbereavement.
32 EXERCISES
This can be done using stsplit:

. stsplit brv, after(time=dosp) at(0)
. recode brv -1=0 0=1
This syntax of stsplit splits the records at the death of spouse (or 1/1/2000 if the
spouse is still alive). The variable brv takes the values 1 for the pre bereavement
part and 0 for the post bereavment part and the recode command changes these to
0 and 1 respectively.
To see the effect on couple 3
. list id sex doe dosp dox brv _t0 _t _d fail if couple==3
We see that, of this couple, only the woman was bereaved during follow-up (it is
impossible for both of a couple to contribute person-time to the bereaved category).
This woman was classified as not bereaved during age 83.87 and 84.41 and bereaved
during ages 84.41 and 84.82. Study the data for the other couples mentioned above.
(d) Now find the (crude) effect of bereavement
. streg brv, dist(exp)
(e) Since there is a strong possibility that the effect of bereavement is not the same for
men as for women, use streg to estimate the effect of bereavement separately for
men and women. Do this both by fitting separate models for males and females (e.g.
streg brv if sex==1) as well as by using a single model with an interaction term
(you may need to create dummy variables). Confirm that the estimates are identical
for these two approaches.
(f) Controlling for age. There is strong confounding by age. Use stsplit to expand
the data by 5 year agebands, and check that the rate is increasing with age. Use
streg to find the effect of bereavement controlled for age. If you wish to study the
distribution of age then it is useful to know that age at entry and exit are stored in
the variables _t0 and _t respectively.
(g) Now estimate the effect of bereavement (controlled for age) separately for each sex.
(h) We have assumed that any effect of bereavement is both immediate and permanent.
This is not realistic and we might wish to improve the analysis by further subdividing
the postbereavement followup. How might you do this? (you are not expected to
actually do it)
(i) Analysis using Cox regression. We can also model these data using Cox regres-
sion. Provided we have stset the data with attained age as the time scale and split
the data (using stsplit) to obtain separate observations for the bereaved and non-
bereaved person-time the following command will estimate the effect of bereavement
adjusted for attained age.
. stcox brv
That is, we do not have to split the data by attained age (although we can fit the
model to data split by attained age and the results will be the same).
(j) Use the Cox model to estimate the effect of bereavement separately for males and
females and compare the estimates to those obtained using Poisson regression.
33
130. Melanoma: Understanding splines

Stata addon required! This exercise requires the Stata user-written command rcsgen.
See Section 2.3 (page 6) for details and installation instructions.
This question is for those who want to understand the calculations made when using splines
and how the constraints enable a smooth function to be fitted. This will be demonstrated
by fitting a Poisson model with no covariates (other than follow-up time). First load the
melanoma data with follow-up to 10 years and split the time scale with intervals of one
month.
. use melanoma
. gen female = sex == 2
. stset surv_mm, failure(status=1,2) scale(12) exit(time 120) id(id)
. stsplit fu, every(=1/12)
. gen risktime = _t - _t0
. collapse (sum) d = _d risktime (min) start=_t0 (max) end=_t, ///
by(fu female year8594 agegrp)
(a) Fit a Poisson model for all cause survival with one parameter for each interval. Predict
the hazard function and plot this against follow-up time.
. egen interval = group(start)
. gen midtime = (start + end)/2
. glm d ibn.interval, family(poisson) link(log) lnoffset(risktime) nocons
// predict the baseline (one parameter for each interval)

. predict haz_grp, nooffset
. replace haz_grp = haz_grp*1000
. twoway (scatter haz_grp midtime) ///
, xtitle("Years from diagnosis") ///
ytitle("Baseline hazard (1000 pys)") ///
ylabel(5 10 20 50 100 150, angle(h)) ///
name(piecewise, replace)
How many parameters have been used to estimate the baseline hazard?
(b) We will now use piecewise linear splines. To simplify things we will only have one
knot at 1.5 years. We will first fit the equivalent of two separate linear functions, one
before the knot and one after the knot. The model is as follows,
ln h(t) = 0 + 1 t + 2 (t > 1.5) + 3 (t 1.5)+
Note the use of the + notation, where u+ = u if u > 0 and 0 otherwise. Write down
the functional form of the linear functions before and after the knot at 1.5 years.
Now fit this model and compare the fitted values to the piecewise estimate in part
130a.
. gen lin_s1 = midtime
. gen lin_int2 = (midtime>1.5)
. gen lin_s2 = (midtime - 1.5)*(midtime>1.5)
34 EXERCISES
// Fit two separate linear regression lines (4 parameters)

. glm d lin_s1 lin_int2 lin_s2 , family(poisson) link(log) lnoffset(risktime)
. predict haz_lin1, nooffset

. replace haz_lin1 = haz_lin1*1000
(line haz_lin1 midtime if midtime<=1, lcolor(red)) ///
(line haz_lin1 midtime if midtime>1, lcolor(red)) ///
xline(1.5, lcolor(black) lpattern(dash)) ///
ylabel(5 10 20 50 100 150, angle(h)) ///
legend(off) ///
name(linear1, replace)
Calculate the intercept and gradient before the knot and after the knot at 1.5 years.
(c) Now we will force the function to be continuous at the knot. This can be done by
dropping the second intercept term. Thus the model is,
ln h(t) = 0 + 1 t + 2 (t 1.5)+
Fit this model and plot the estimated hazard against the piecewise estimates.
// Force the functions to join at the knot (3 parameters)
. glm d lin_s1 lin_s2 , family(poisson) link(log) lnoffset(risktime)
. predict haz_lin2, nooffset

. replace haz_lin2 = haz_lin2*1000
(line haz_lin2 midtime, lcolor(red)) ///
xline(1.5, lcolor(black) lpattern(dash)) ///
ylabel(5 10 20 50 100 150, angle(h)) ///
legend(off) ///
name(linear2, replace)
Calculate the gradient before the knot and after the knot at 1.5 years.
(d) We will now perform a similar exercise for cubic splines. Again we will have a single
knot, this time at 2 years. We will fit the equivalent two separate cubic functions, one
before the knot and one after the knot, and then start to introduce the constraints
that ensure the function is smooth. There will be eight parameters in the model (3
polynomial terms and an intercept for each of the two intervals).
3
X 3
X
ln h(t) = k tk + k+4 (t 2)k+
k=0 k=0
35
Fit this model, predict the hazard and plot, comparing the fit to the piecewise esti-
mates.
. gen cubic_s1 = midtime
. gen cubic_s2 = midtime^2
. gen cubic_s3 = midtime^3
. gen cubic_int = midtime>2
. gen cubic_lin = (midtime - 2)*(midtime>2)
. gen cubic_quad = ((midtime - 2)^2)*(midtime>2)
. gen cubic_s4 = ((midtime - 2)^3)*(midtime>2)
. glm d cubic* , family(poisson) link(log) lnoffset(risktime)

. predict haz_cubic1, nooffset
. replace haz_cubic1 = haz_cubic1*1000
(line haz_cubic1 midtime if midtime<=2, lcolor(red)) ///
(line haz_cubic1 midtime if midtime>2, lcolor(red)) ///
xline(2, lcolor(black) lpattern(dash)) ///
ylabel(5 10 20 50 100 150, angle(h)) ///
legend(off) ///
name(cubic1, replace)
(e) We will now constrain the function to be continuous at the knot. This can be done
by droping the second intercept term. The model becomes,
3
X 3
X
ln h(t) = k tk + k+3 (t 2)k+
k=0 k=1
(The subscript for the second sum, starts at k = 1 rather than k = 0.

Fit this model by excluding the variable cubic int from the model. Plot the predicted
hazard to ensure that it is continuous at the knot. Explain why the function does
not look smooth.
(f) Now we will force the first derivative to be continuous. This can be done by dropping
the second linear term from the model.
3
X 3
X
ln h(t) = k tk + k+2 (t 2)k+
k=0 k=2
Fit this model by excluding the variable cubic lin from the model. Plot the predicted
hazard to ensure that it is continuous at the knot. Does the function look smoother?
(g) Finally we will force the second derivative to also be continuous. This can be done
by dropping the second quadratic term from the model.
3
X
ln h(t) = k tk + 4 (t 2)3+
k=0
Fit this model by excluding the variable cubic quad from the model. Plot the pre-
dicted hazard. Compare the fits of the models with the different constraints.
36 EXERCISES
(h) For the models we fit we usually use restricted cubic splines. These are constrained
to be linear before the first knot and after the final knot. We nearly always put the
boundary knots at the minimum and maximum event times and so the restriction of
linearity is not actually within the range of our data, but the linear restrictions help
to stabalise the estimated function. A good derivation of how the linear restrictions
are imposed can be found in Appendix B of Royston and Parmer (2002) [7].
Generate the restricted cubic spline basis functions with 4 degrees of freedom (5
knots). As we have collapsed data we need to use the fw(d) (frequency weights)
option as we place the knots evenly acording to the distribution of events times, i.e.
in this case at the 0th (minimum), 25th , 50th , 75th and 100th (maximum) centiles of
the distribution of event times.
. rcsgen midtime, gen(rcs) df(4) fw(d)
. global knots r(knots)
We have stored the location of the knots in a global macro, so we can add them to
later plots.
(i) The first spline variable, rcs1, is just copy of our x variable, midtime. Fit a model
where you assume that the log hazard function is a linear function of log time and
plot the fitted function.
. glm d rcs1, family(poisson) link(log) lnoffset(risktime)
. estimates store rcs1
. predict haz_rcs1, nooffset
. replace haz_rcs1 = haz_rcs1*1000
(line haz_rcs1 midtime, lcolor(red)) ///
ylabel(5 10 20 50 100 150, angle(h)) ///
legend(off) ///
name(rcs1, replace)
Does this model look a good fit?

(j) Now add the remaining spline variables, rcs2-rcs4, to the model and perform a
likelihood ratio test to see if there is evidence of non linearity.
. glm d rcs*, family(poisson) link(log) lnoffset(risktime)
. estimates store rcs2
. lrtest rcs1 rcs2
Plot the fitted function against the piecewise estimates and show the location of the
knots as reference lines.
xline(\$knots, lcolor(black) lpattern(dash)) ///
ylabel(5 10 20 50 100 150, angle(h)) ///
legend(off) ///
name(rcs2, replace)
37
(k) We will now show the restriction of linearity beyond the boundary knots by moving
them within the range of the data. Recalculate the restricted cubic splines with knots
at 1, 2 and 3 years. Plot the estimated hazard function.
. drop rcs*
. rcsgen midtime, gen(rcs) knots(1 2 3) fw(d)
. global knots r(knots)
. glm d rcs*, family(poisson) link(log) lnoffset(risktime)
xline(\$knots , lcolor(black) lpattern(dash)) ///
ylabel(5 10 20 50 100 150, angle(h)) ///
legend(off) ///
name(rcs3, replace)
38 EXERCISES
131. Modelling cause-specific mortality using flexible parametric models
Stata addon required! This exercise requires the Stata user-written command stpm2.
We will now fit some models with the linear predictor on the log cumulative hazard scale
using flexible parametric survival models (Royston-Parmar models).
Load amd stset the Melanoma data.

. stset surv_mm, failure(status==1) exit(time 120.5) scale(12)
(a) Plot a Kaplan-Meier curve for the study population as a whole.

sts graph
(b) Fit a Weibull model using stpm2. In a Weibull model the log cumulative hazard
function is a linear function of log(time). This can be fitted with the scale(hazard)
and df(1) options.
stpm2, scale(hazard) df(1)
predict s1, surv
predict h1, hazard
The two prediction commands will estimate the survival and and hazard functions
respectively.
Overlay the estimated survival function from a Weibull model with the Kaplan-Meier
curve.
sts graph, addplot(line s1 _t, sort) name(km1, replace)
Does the Weibull model fit well? What assumptions are made about the shape of the
hazard function with a Weibull model? Is it possible to assess this from looking at
the survival curve?
(c) In Stata we can obtain an estimate of the hazard function using weighted kernel-
density estimation using sts graph, hazard. I usually use the kernel(epan2) op-
tion as this generally works better than the default.
Plot this hazard estimate, overlaying the the estimated hazard from the Weibull
model.
sts graph, hazard kernel(epan2) addplot(line h1 _t, sort) name(hazard1, replace)
You should now understand why the Weibull model does not fit very well.
(d) We will now relax the assumption of linearity of the log cumulative hazard with
respect to log time by incorporating restricted cubic splines into the model. We will
initially use 4 df (5 knots) using the default knot locations.
stpm2, scale(hazard) df(4)
predict s4, surv
predict h4, hazard
Now add the predictions of the survival and hazard functions to the non-parametric
survival and hazard functions.
39
sts graph, addplot(line s4 _t, sort) name(km4, replace) ///

xline(e(bhknots) e(boundary_knots))
sts graph, hazard kernel(epan2) addplot(line h4 _t, sort) name(hazard4, replace) ///
line(e(bhknots) e(boundary_knots))
Are the fitted curves better than the Weibull model?

(e) Fit a Cox model with the diagnosis period (year8594) as the only covariate.
. stcox year8594
(f) Fit the equivalent flexible parametric survival model on the log cumulative hazard
scale with 4 degrees of freedom for the baseline.
. stpm2 year8594, scale(hazard) df(4) eform
Compare the estimated hazard ratio, 95% confidence interval and statistical signifi-
cance to the Cox model.
(g) Obtain predicted values of the survival and hazard functions and plot these functions
by calendar period of diagnosis
. predict s1ph, survival

. predict h1ph, hazard per(1000)
. twoway (line s1ph _t if year8594 == 0, sort) ///
(line s1ph _t if year8594 == 1, sort) ///
, legend(order(1 "1975-1984" 2 "1985-1994") ring(0) pos(1) col(1)) ///
xtitle("Time since diagnosis (years)") ///
ytitle("Survival")
. twoway (line h1ph _t if year8594 == 0, sort) ///

(line h1ph _t if year8594 == 1, sort) ///
, legend(order(1 "1975-1984" 2 "1985-1994") ring(0) pos(1) col(1)) ///
ytitle("Cause specific mortality rate (per 1000 pys)")
(h) Add the option yscale(log) to the hazard plot to display the hazard function on the
log scale. Why is the difference between the two lines constant over the time scale?
(i) Note that there are 4 rcs terms because of the df(4) option. We can investigate
more or less degrees of freedom for the baseline. It is easiest to do this in a loop. We
will also store the model estimates using estimates store and predict the baseline
survival and hazard functions.
forvalues i = 1/6 {
stpm2 year8594, scale(hazard) df(i) eform
estimates store dfi
predict h_dfi, hazard per(1000) zeros
predict s_dfi, survival zeros
}
Compare the hazard ratios, AIC and BIC from the different models.
. estimates table df*, eq(1) keep(year8594) se stats(AIC BIC)
According to the AIC and BIC how many degrees of freedom should be used for the
baseline? Does it matter for the interpretation of the estimated hazard ratio?
About AIC and BIC AIC (Akaike information criterion) and BIC (Bayesian infor-
mation criterion) are two popular measures for comparing the relative goodness-of-fit
40 EXERCISES
of statistical models. The AIC and BIC are defined as:
AIC = 2 ln(likelihood) + 2k
BIC = 2 ln(likelihood) + ln(N )k
where k = number of parameters estimated and N = number of observations.
Given a set of candidate models for the data, the preferred model is the one with
the minimum AIC/BIC value. Hence, the measures not only reward goodness of fit,
but also include a penalty that is an increasing function of the number of estimated
parameters. AIC uses a fixed constant, 2, in the penalty term whereas the penalty
in BIC is a function of the number of observations. It is not always obvious how
number of observations should be defined for time-to-event data, particularly for
grouped or split data. Volinsky and Raftery (2000) [8] suggest using the number of
events for N in the BIC penalty term for survival models. The estimates stats
command contains an option n(#) for specifying N .
In many circumstances both the AIC and BIC will suggest the same model. For
population-based survival data, the number of observations is large so BIC will pe-
nalize models with additional parameters more strongly than AIC.
(j) Compare the estimated baseline survival and hazard functions for the models with
varying degrees of freedom.
. line s_df* _t, sort ///
legend(ring(0) cols(1) pos(1)) ///
ytitle("Survival") ///
name(compsurv, replace)
. line h_df* _t, sort ///

ytitle("Cause specific mortality rate (per 1000 pys)") ///
name(comphazard, replace)
Comment on the agreement between the different choices of degrees of freedom.

(k) The previous question compares using a different number of knots placed at their
default locations. However, it is also reasonable to ask whether placing the knots in
different places leads to different model fits. Run the code in the solution Do file to
fit 10 different models with 5df (6 knots) with the 4 internal knots placed at random
centiles of the distribution of event times. Dont worry about understanding the code
(unless you want to).
Comment on the agreement in the hazard ratios, and the baseline survival and hazard
functions between the different models.
(l) Now include sex and age (in categories) in the model. First fit a Cox model and
compare the parameter estimates
. stcox female year8594 i.agegrp
. estimate store cox
. stpm2 i.sex year8594 i.agegrp, df(4) scale(hazard) eform
. estimates store stpm2_ph
41
(m) Explain why the estimates from the Cox model and the flexible parametric model are
so similar.
(n) As models become more complex, we may want predictions for specific combina-
tions of covariates. This is particularly the case when be have continuous covariates.
stpm2s predict command has useful at() and zeros options. The at() option
requests predictions at specified values of covariates and the zeros options requests
that any variables not listed in the at() option are set to zero.
Another useful option is the timevar(varname) option. This requests that the pre-
dictions are at values of time specified in varname rather than the default t. This
is useful for very large datasets or when wanting to predict survival for various com-
binations of covariates at one specific time, e.g. 5 years.
i. Define a new variable, temptime with 200 observations taking values of time
between 0 and 10 and predict the baseline survival function.
. estimates restore ph
. range temptime 0 10 200
. predict S0, survival zeros timevar(temptime)
. line S0 temptime, sort
What combination of covariates does the baseline survival represent?
ii. Using the at() and zeros options predict the hazard function for females aged
75+ diagnosed in 1985-1994 for values of temptime with a 95% confidence inter-
val.
. predict S0, survival zeros timevar(temptime)
. line S0 temptime, sort
. predict S_F_8594_age75, survival ///

at(female 1 year8594 1 agegrp 3) timevar(temptime) ci
. twoway (rarea S_F_8594_age75_lci S_F_8594_age75_uci temptime, pstyle(ci)) ///

(line S_F_8594_age75 temptime) ///
, legend(off) ///
ytitle("S(t)") ///
title("Female age 75+ diagnosed 1985-1994")
42 EXERCISES
132. Modelling time-dependent effects using flexible parametric models
We will now fit a model where the effect of age group is time-dependent. Reload and
stset the data.

. stset surv_mm, failure(status==1) exit(time 60.5) scale(12)
We have restricted follow-up to five years.
(a) First we will fit a Cox model and assess the proportional hazards assumption using
Schoenfeld residuals. One can obtain a plot of the scaled Schoenfeld residuals, with
a smoother, for a chosen predictor using the following command.
. estat phtest, plot(3.agegrp)
We will use a loop to produce plots for each agegrp and add a horizontal line at the
value of the estimated log hazard ratio.
. stcox female year8594 i.agegrp,
. forvalue i = 1/3 {
local beta = _b[i.agegrp]
estat phtest, plot(i.agegrp) name(sch_agei, replace) ///
yline(0 beta) msize(small) msymbol(Oh) bw(0.4)
}
. estat phtest, detail
Is the effect of age proportional?

(b) Now fit a flexible parametric proportional hazards model with 4 df for the baseline.
Note that when we go on to use the tvc() option you cant use Statas factor variables,
so we create dummy variables for age group.
. tab agegrp, gen(agegrp)
. stpm2 female year8594 agegrp2-agegrp4, df(4) scale(hazard) eform
. estimates store ph
Predict and plot the hazard function for each age group for males diagnosed in 1975-
1984. Note the use of the at() and zeros options.
. predict h_age1, hazard zeros per(1000)
. predict h_age2, hazard at(agegrp2 1) zeros per(1000)
. twoway (line h_age1 _t, sort) ///

(line h_age2 _t, sort) ///
,xtitle("Time since diagnosis (years)") ///
legend(order(1 "<45" 2 "45-59" 3 "60-74" 4 "75+") ring(0) pos(1) cols(1))
43
(c) Now fit a model with time-dependent effects for age group. We will do this using 2
degrees of freedom for each age category.
. stpm2 female year8594 agegrp2-agegrp4, df(4) scale(hazard) ///
tvc(agegrp2 agegrp3 agegrp4) dftvc(2)
. estimates store nonph
Perform a likelihood ratio test comparing the proportional hazards model with the
non-proportional hazards (for age) model. Is there evidence of a non-proportional
effect?
. lrtest ph nonph
(d) Now predict the hazard function for each age group. Note that the prediction com-
mand is identical to that used in the proportional hazards model as stpm2 knows
which variables have time-dependent effects and takes this into account when pre-
dicting.
. predict h_age1_tvc, hazard zeros per(1000)
. predict h_age2_tvc, hazard at(agegrp2 1) zeros per(1000)
. twoway (line h_age1 h_age1_tvc _t, sort lcolor(red red) lpattern(solid dash)) ///
(line h_age2 h_age2_tvc _t, sort lcolor(blue blue) lpattern(solid dash)) ///
(line h_age3 h_age3_tvc _t, sort lcolor(magenta magenta) lpattern(solid dash)) ///
(line h_age4 h_age4_tvc _t, sort lcolor(green green) lpattern(solid dash)) ///
,xtitle("Time since diagnosis (years)") ///
legend(order(1 "<45" 2 "45-59" 3 "60-74" 4 "75+") ring(0) pos(1) cols(1)) ///
name(hazard_tvc, replace)
(e) Obtain a prediction of the hazard ratio as a function of time for each age group.
. predict hr2, hrnumerator(agegrp2 1) ci
Note that by default the hrdenominator option sets all covariates to zero. As we
only have one covariate with a time-dependent effect we can leave this unspecified.
Plot these hazard ratios versus follow-up time on the same graph. What happens to
the hazard ratios as follow-up time increases? Also plot the hazard ratio for the oldest
group with a 95% confidence interval. Explain why the hazard ratio for the oldest
age group is so high early on in the time-scale (hint: look at the baseline hazard)
. twoway (line hr2 hr3 hr4 _t, sort), ///
yscale(log) ylabel(1 2 10 20 50) ///
legend(order(1 "Age 45-59" 2 "Age 60-74" 3 "Age 75+") ring(0) pos(1) cols(1)) ///
ytitle("Hazard ratio") ///
name(hr, replace)
. twoway (rarea hr4_lci hr4_uci _t, sort pstyle(ci)) ///

(line hr4 _t, sort) ///
,legend(off) yscale(log) ylabel(1 2 10 20 50) ///
ytitle("Hazard ratio") ///
name("hr_age4", replace)
44 EXERCISES
(f) Obtain and plot with 95% confidence intervals the difference in the hazard rates
between the oldest and youngest age groups for males in 1975-1984.
. predict hdiff4, hdiff1(agegrp4 1) ci per(1000)

. twoway (rarea hdiff4_lci hdiff4_uci _t, sort) ///
(line hdiff4 _t, sort) ///
,legend(off) ///
ytitle("Difference in hazard rate") ///
name(hdiff, replace)
Explain why the hazard difference is small early on in the time-scale, when the hazard
ratio is at is greatest.
(g) Predict and plot the survival function for the youngest and oldest age groups for
females diagnosed in 1985-1994.
. predict s1, surv at(female 1 year8594 1) zeros

. predict s2, surv at(agegrp4 1 female 1 year8594 1) zeros
Obtain and plot with 95% confidence intervals the difference in the survival functions
between the oldest and youngest age groups for females diagnosed in 1985-1994.
. predict sdiff4, sdiff1(agegrp4 1 sex 2 year8594 1) ///

sdiff2(agegrp4 0 sex 2 year8594 1) ci
. twoway (rarea sdiff4_lci sdiff4_uci _t, sort) ///
(line sdiff4 _t, sort) ///
,legend(off) ///
ytitle("Difference in survival functions") ///
name(sdiff, replace)
(h) Fit models with 1, 2 and 3 df for the time-dependent effect of age. Use the AIC
and BIC to compare models. Compare the estimated time-dependent hazard ratio
for the oldest age group compared to the youngest (also compare the 95% confidence
intervals). You may want to exclude the first month from you plot as the hazard
ratio is very high close to zero.
forvalues i = 1/3 {
stpm2 i.sex year8594 agegrp2-agegrp4, df(4) scale(hazard) ///
tvc(agegrp2 agegrp3 agegrp4) dftvc(i)
estimates store dftvci
predict hr4_dfi, hrnumerator(agegrp4 1) ci
}
count if _d==1
estimates stats dftvc*, n(r(N))
twoway (line hr4_df1 hr4_df1_lci hr4_df1_uci _t, sort lcolor(red..) ///

lpattern(solid dash dash) lwidth(medthick thin thin)) ///
(line hr4_df2 hr4_df2_lci hr4_df2_uci _t, sort lcolor(midblue..) ///
(line hr4_df3 hr4_df3_lci hr4_df3_uci _t, sort lcolor(midgreen..) ///
if _t>0.1, ///
yscale(log) ///
ylabel(1 2 4 8 20 50, angle(h)) ///
45
legend(order(1 "1 df" 4 "2 df" 7 "3 df") ring(0) pos(1) cols(1)) ///
ytitle("Hazard Ratio") ///
yscale(log) ///
name(tvc_df_comp, replace)
(i) We will now also let the effect of sex be time-dependent to illustrate when there
are two time-dependent effects the hazard ratios are not the exactly the same. Add
female to the tvc option.
. stpm2 female agegrp2-agegrp4, df(4) scale(hazard) ///
tvc(agegrp2 agegrp3 agegrp4 female) dftvc(3)
Use the hrnumerator and hrdenominator options of the predict command to obtain
a prediction of the hazard ratio for female for the youngest and the oldest age groups.
Add the ci option to obtain confidence intervals. Compare the resulting curves and
their 95% confidence intervals to show that the curves are similar, but not identical.
. predict hr_f_age1, hrnum(female 1) ci
. predict hr_f_age4, hrnum(female 1 agegrp4 1) hrdenom(agegrp4 1) ci
. twoway (line hr_f_age1* hr_f_age4* _t if _t>0.1, sort yscale(log))
(j) additional topic - not covered in lectures If we were modelling on the log hazard
scale, then including exactly the same covariates and time-dependent effect, the haz-
ard ratio would be equivalent. Such a model can be fitted using the strcs command.
This command is much slower than stpm2 as it requires numerical integration (using
Gauss-Legendre quadrature) to estimate the parameters.
. strcs female agegrp2-agegrp4, df(4) ///
tvc(agegrp2 agegrp3 agegrp4 female) dftvc(3) nodes(50)
. predict hr_f_age1b, hrnum(female 1) ci
. predict hr_f_age4b, hrnum(female 1 agegrp4 1) hrdenom(agegrp4 1) ci
. twoway (line hr_f_age1b* hr_f_age4b* _t if _t>0.1, sort yscale(log))

46 EXERCISES
133. Modelling cause-specific mortality on other scales (proportional odds and Aranda-
Ordaz link function) using stpm2
This question uses the Melanoma data. Load and stset the data.
use melanoma, clear

gen female = sex == 2
stset surv_mm, failure(status=1) scale(12) exit(time 60.5)
(a) Fit a proportional hazards model to the melanoma data with age group, sex and
calendar year as covariates. Predict the survival and hazard functions for the youngest
and oldest age groups for those diagnosed in 1975-1984. Store the model estimates.
stpm2 female i.agegrp year8594, scale(hazard) df(4) eform
forvalues i = 0/3 {
predict s_agei_ph, surv at(agegrp i) zeros
predict h_agei_ph, hazard at(agegrp i) zeros
}
estimates store ph
(b) Now fit a proportional odds model and predict the survival and hazard functions.
You just need to to change the scale(hazard) option to scale(odds)
stpm2 female i.agegrp year8594, scale(odds) df(4) eform
forvalues i = 0/3 {
predict s_agei_po, surv at(agegrp i) zeros
predict h_agei_po, hazard at(agegrp i) zeros
}
estimates store po
Interpret the effect of the covariate female

(c) Compare the predict survival and hazard function between the proportional odds and
proportional hazards models. Explain why they are not the same.
twoway (line s_age0_ph _t, sort) ///
(line s_age0_po _t, sort) ///
(line s_age3_ph _t, sort) ///
(line s_age3_po _t, sort) ///
, name(survcomp, replace)
twoway (line h_age0_ph _t, sort) ///

(line h_age0_po _t, sort) ///
(line h_age3_ph _t, sort) ///
(line h_age3_po _t, sort) ///
, name(hazcomp,replace)
(d) According to the BIC and AIC, which is the best fitting model?
count if _d == 1
estimates stats ph po, n(r(N))
47
(e) For the proportional odds model the hazards will not be proportional. Predict and
plot the hazard ratio for females in the youngest age group diagnosed in 1975-1984.
predict hr_female_age0_7584, hrnum(female 1) hrdenom(female 0) ci
twoway (rarea hr_female_age0_7584_lci hr_female_age0_7584_uci _t, sort pstyle(ci)) ///
(line hr_female_age0_7584 _t, sort) ///
,legend(off) ///
xtitle("Years since diagnosis") ///
ytitle("Hazard Ratio") ///
title("HR for sex (age<45, diagnoised 1975-1984)") ///
name(HR1, replace)
(f) The hazard ratio for females will be different at different levels of other covariates.
Show this by now calculating the hazard ratio for females in the oldest age group
diagnosed in 1975-1984.
predict hr_female_age3_7584, hrnum(female 1 agegrp 3) hrdenom(female 0 agegrp 3) ci
twoway (line hr_female_age0_7584 _t, sort) ///
(line hr_female_age3_7584 _t, sort) ///
,name(HR2, replace)
(g) Now fit a model using the Aranda-Ordaz link function using the scale(theta) option.
Compare the AIC/BIC with the proportional hazard and proportional odds model.
stpm2 female i.agegrp year8594, scale(theta) df(4)
estimates store ao
count if _d == 1
estimates stats ph po ao, n(r(N))
(h) The proportional odds model provides a better fit. Calculate the estimated value of
with 95% confidence intervals. Explain why this is the case.
lincom [ln_theta][_cons], eform
48 EXERCISES
140. Probability of death in a competing risks framework (cause-specific survival)
Stata addon required! This exercise requires the Stata user-written commands, stpm2,
stcompet, stpepemori stcompadj and stpm2cif. See Section 2.3 (page 6) for details and
installation instructions.
This question gives an introduction to some of the methods available for competing risks
analyses. We will be estimating similar quantities to 250 and 251 but we are now working
in a cause-specific survival framework. To carry out the exercises you will need to install
some user-written commands from within Stata. The stcompet command estimates the
cumulative incidence function (CIF) non-parametrically. The stpm2cif command esti-
mates the CIF through post-estimation after fitting a flexible parametric model.
(a) Load the colon data dropping those with missing stage.
use colon, clear
drop if stage ==0
gen female = sex==2
If you summarize status you will notice that there are deaths from both cancer and
other causes. Plot the complement of the Kaplan-Meier estimate for males (i.e., 1
minus Kaplan-Meier survival estimate) for both cancer and other causes. Describe
what you see.
A common confusion when competing risks are present is to think that the probability
of death from cancer can be obtained by taking the complement of the Kaplan-Meier
estimate (1-KM). By doing this we treat deaths from other causes as censored. If we
can assume independence, that is the patients dying from other causes would have
been at no systematically higher or lower risk of dying from cancer, then we estimate
the marginal probability of death, i.e., the probability of death in the hypothetical
world where it is not possible to die of other causes.
The appropriate estimate for the real world probability of death from cancer when
competing risks are present is the cumulative incidence function. That is the propor-
tion of patients that have died from cancer at a certain time in the follow-up period
taking into account competing causes of death.
(b) Use the stcompet command to estimate the cumulative incidence function for both
cancer and other causes. Plot the cumulative incidence functions for males along with
the complements of the Kaplan-Meier estimates from part (a). What do you notice?
stset surv_mm, failure(status==1) scale(12) exit(time 120.5)
stcompet CIF_sex=ci, compet1(2) by(sex)
gen CIF_sex_cancer=CIF_sex if status==1
gen CIF_sex_other=CIF_sex if status==2
49
(c) Obtain estimates of the CIF for cancer and other causes by age group. Plot and
interpret the curves.
stcompet CIF_age=ci, compet1(2) by(agegrp)
twoway (line CIF_age _t if agegrp == 0 & status == 1, sort connect(stepstair)) ///

(line CIF_age _t if agegrp == 1 & status == 1, sort connect(stepstair)) ///
, legend(order(1 "<45" 2 "45-59" 3 "60-74" 4 "75+") ring(0) pos(5) cols(1)) ///
ytitle("CIF") ///
title("Cancer") ///
name(CIF_age1,replace)
twoway (line CIF_age _t if agegrp == 0 & status == 2, sort connect(stepstair)) ///

, legend(order(1 "<45" 2 "45-59" 3 "60-74" 4 "75+") ring(0) pos(11) cols(1)) ///
ytitle("CIF") ///
title("Other causes") ///
name(CIF_age2,replace)
graph combine CIF_age1 CIF_age2, nocopies ycommon
(d) Now obtain the CIF for cancer and other causes by stage group. Plot the results.
Explain why those diagnosed with regional and distant stage are less likely to die
from other causes when compared to those with localized disease.
. stcompet CIF_stage=ci, compet1(2) by(stage)
. twoway (line CIF_stage _t if stage == 1 & status == 1, sort connect(stepstair)) ///

(line CIF_stage _t if stage == 2 & status == 1, sort connect(stepstair)) ///
, legend(order(1 "local" 2 "regional" 3 "distant") ring(0) pos(5) cols(1)) ///
ytitle("CIF") ///
title("Cancer") ///
name(CIF_stage1,replace)
. twoway (line CIF_stage _t if stage == 1 & status == 2, sort connect(stepstair)) ///

, legend(order(1 "local" 2 "regional" 3 "distant") ring(0) pos(1) cols(1)) ///
ytitle("CIF") ///
title("Other causes") ///
name(CIF_stage2,replace)
. graph combine CIF_stage1 CIF_stage2, nocopies ycommon

50 EXERCISES
(e) We will now fit a Fine and Gray model with death due to cancer as the main outcome
of interest, and death from other causes as the competing event.
i. Fit a model that includes sex as a covariate.
stcrreg i.sex, compete(status == 2)
How would you interpret the estimated effect of sex (SHR)?
ii. Based on the results from the fitted model, plot the cancer-specific CIF for males
and females, respectively, and test if there is evidence of a difference by sex.
predict cif_males, basecif
gen cif_females = 1 - (1-cif_males)êxp(_b[2.sex])
graph twoway line cif_males cif_females _t, ///

sort connect(step step) yscale(range(0 0.6)) ylabel(0(0.2)0.6) ///
legend(order(1 "Males" 2 "Females")) ///
ytitle(Cause-specific cumulative incidence) ///
xtitle(Time since diagnosis (years))
stpepemori sex, compet(2)
iii. The cause-specific CIFs can also be retrieved via the stcurve function in Stata.
Using, stcurve, re-plot the cancer-secific CIFs by sex and verify that the esti-
mates are the same by plotting them side by side the estimates that you calculated
from first principles.
stcurve, cif at1(sex = 1) at2(sex=2)
(f) Now, fit another Fine and Gray model but this time with death due to other causes as
the main event of interest, and death due to cancer as the competing event. Interpret
the parameter estimate and state you conclusion about the effect of sex based on the
model output.
(g) We will now fit a competing risks model using the flexible parametric approach. In
order to do this we will first need to expand the data set so that each patient has two
rows of data - one for each cause of death.
i. Expand the data and have a look at the new data set.
. expand 2
. bysort id: gen cause=_n
. gen cancer=(cause==1)
. gen other=(cause==2)
. gen event=(cause==status)
. list id status cause sex event in 1/8, sepby(id)
51
ii. Fit a flexible parametric model for cancer and other causes simultaneously. In-
clude sex as a covariate assuming that the effect of sex is the same for both cancer
and other causes. Interpret the effect of sex.
. stset surv_mm, failure(event) scale(12)
. stpm2 cancer other female, scale(hazard) ///
rcsbaseoff dftvc(3) nocons tvc(cancer other) eform nolog
By including the two cause indicators (cancer and other) as both main effects
and time-dependent effects (using tvc option) we have fitted a stratified model
with two separate baselines, one for each cause. For this reason we have used the
rcsbaseoff option together with the nocons option which excludes the baseline
hazard from the model.
iii. We usually would like to allow the effect of covariates to be different for the
different causes. Now fit a model where the effect of sex is allowed to be different
for cancer and other causes.
. gen fem_can = female*cancer
. gen fem_other = female*other
. stpm2 cancer other fem_can fem_other, scale(hazard) ///
Test(using a Wald test) where there is evidence that the effect of sex differs
between cancer and other causes.
. test fem_can = fem_other
iv. Explain how the parameter estimates for the effect of sex on cancer mortality
and other cause mortality are conceptually different from the ones you estimated
in the Fine and Gray models.
(h) Use the stpm2cif postestimation command to obtain the cumulative incidence func-
tions for cancer and other causes for each sex. You will need to run this command
twice - once for each sex. Notice that a new time variable is generated called newt.
You will need to use this instead of t in your plots. Do the results look the same as
the empirical estimates (overlay these to make the comparison easier).
. stpm2cif cancermale othermale, cause1(cancer 1) ///
cause2(other 1)
. stpm2cif cancerfemale otherfemale, cause1(cancer 1 fem_can 1) ///
cause2(other 1 fem_other 1)
See the do file for the relevant Stata code. What are potential reasons for any dis-
agreement between the empirical and model estimates?
(i) Think about an alternative way to present the results. Try stacking the cumulative
incidence functions for cancer and other causes (see do file for code to plot this).
(j) So far the model only included one covariate (male/female). We now want to adjust
the model for age. However, we dont believe that the effect of age is the same for
both cancer and other causes of death.
i. To allow the effect of age to vary for the two causes create interaction terms
between age group and the causes of death.
. forvalues i = 0/3 {
gen ageican=(agegrp==i & cancer==1)
gen ageioth=(agegrp==i & other==1)
}
52 EXERCISES
ii. Fit a flexible parametric model including sex and the interaction terms between
age group and cause.
. stpm2 cancer other fem_can fem_oth ///
age1ca age2ca age3ca age1oth age2oth age3oth , scale(hazard) ///
Interpret the hazard ratios
(k) Predict the cause-specific CIFs for males in the youngest and oldest age groups.
. stpm2cif cancermale_age0 othermale_age0, cause1(cancer 1) ///
cause2(other 1)
. stpm2cif cancermale_age3 othermale_age3, cause1(cancer 1 age3can 1) ///
cause2(other 1 age3oth 1)
(l) Now incorporate sex and age as time-dependent effects for cancer in the model (use 3
df). Estimate the CIFs and compare to the model that assumes proportional hazards.
age1can age2can age3can age1oth age2oth age3oth , scale(hazard) ///
rcsbaseoff dftvc(cancer:4 other:4 3) nocons ///
tvc(cancer other fem_can age1can age2can age3can) eform nolog
. stpm2cif cancermale_age0_tvc othermale_age0_tvc, cause1(cancer 1) ///

cause2(other 1)
. stpm2cif cancermale_age3_tvc othermale_age3_tvc, cause1(cancer 1 age3can 1) ///
(m) OPTIONAL EXERCISE By default stpm2 defines knot positions for all events
and does not distinguish between events types. We will now refit the model, but use
the default knot positions obtained when fitting each cause separately.
i. Produce a histogram of the event times separately by event status. Is the distri-
bution of events similar for each cause?
. hist _t if _d==1, by(status)
ii. Fit separate models for each cause and store the knot locations. As we are fitting
time-dependent effects for sex and age for cancer, we will include these in the
model.
. stpm2 fem_can age1can age2can age3can if cancer == 1, ///
df(4) scale(hazard) dftvc(3) ///
tvc(fem_can age1can age2can age3can) eform nolog
. global knots_cancer e(bhknots)
. global knots_cancer_tvc e(tvcknots_age1can)
. stpm2 fem_oth age1oth age2oth age3oth if other == 1, ///

df(4) scale(hazard) eform nolog
. global knots_other e(bhknots)
iii. Refit the model using these knot locations. Has it made any difference to the
estimated CIFs?
age1can age2can age3can age1oth age2oth age3oth , scale(hazard) ///
rcsbaseoff nocons ///
tvc(cancer other fem_can age1can age2can age3can) eform nolog ///
knotstvc(cancer $knots_cancer other $knots_other ///
fem_can $knots_cancer_tvc ///
53
age1can $knots_cancer_tvc ///

age2can $knots_cancer_tvc ///
age3can $knots_cancer_tvc)
. stpm2cif cancermale_age0_tvc2 othermale_age0_tvc2, cause1(cancer 1) ///
cause2(other 1)
. stpm2cif cancermale_age3_tvc2 othermale_age3_tvc2, cause1(cancer 1 age3can 1) ///
(n) We will now estimate cause-specific CIFs in a Cox regression framework using the
stcompadj command. First we need to reload the colon cancer data as the stcompadj
uses as input the data in its original format. The data will be expanded for us in the
background (and we will see later how it is possible to save a copy of the expanded
data that can be used for testing or viewing the results from the Cox regression etc.)
i. Read in the data and stset it with cancer as the main outcome of interest.
use colon, clear

drop if stage ==0
gen female = sex==2
ii. Estimate the cause-specific CIFs by sex from a Cox model that assumes that the
effect of sex on mortality from cancer and other causes, respectively is the same.
Plot the CIFs for males and females in separate graphs.
stcompadj sex=1 , compet(2) gen(Main_males Compet_males)
stcompadj sex=2 , compet(2) gen(Main_females Compet_females)
graph twoway line Main_males Compet_males _t, ///

sort connect(step step) yscale(range(0 1)) ylabel(0(0.2)1) ///
ytitle(Probability of Death) xtitle(Time Since Diagnosis (Years)) ///
title(Males) ///
legend(order(1 "Cancer" 2 "Other")) ///
name(Cox_males, replace)
graph twoway line Main_females Compet_females _t, ///

sort connect(step step) yscale(range(0 1)) ylabel(0(0.2)1) ///
title(Females) ///
name(Cox_females, replace)
graph combine Cox_males Cox_females, ycommon
(o) Weve seen in the earlier exercises that the effect of sex is not the same for both
outcomes under investigation. Estimate another set of CIFs from a model where the
assumption of a shared effect of sex between the outcomes is relaxed. Check how
much the new estimates differ from those that you estimated in the previous exercise
(Overlay the new estimates to make comparisons easier).
stcompadj sex=1 , compet(2) maineffect(sex) competeffect(sex) ///

gen(Main_males_2 Compet_males_2)
stcompadj sex=2 , compet(2) maineffect(sex) competeffect(sex)///
gen(Main_females_2 Compet_females_2)
54 EXERCISES
graph twoway (line Main_males Compet_males _t, lpattern(dash dash) ///

lcolor(navy red) sort connect(step step)) ///
(line Main_males_2 Compet_males_2 _t, lpattern(solid solid) ///
lcolor(navy red) sort connect(step step)), ///
yscale(range(0 0.6)) ylabel(0(0.2)0.6) ///
title(Males) ///
name(Cox_males_2, replace)
graph twoway (line Main_females Compet_females _t, lpattern(dash dash) ///

lcolor(navy red) sort connect(step step)) ///
(line Main_females_2 Compet_females_2 _t, lpattern(solid solid) ///
lcolor(navy red) sort connect(step step)), ///
yscale(range(0 0.6)) ylabel(0(0.2)0.6) ///
title(Females) ///
name(Cox_females_2, replace)
graph combine Cox_males_2 Cox_females_2, ycommon
(p) To test if the effect of sex differs between cancer mortality and other cause mortality
we can fit a Cox regression model that includes an interaction between sex and the
variable that indicates which outcome is being modelled (stcompadj creates a variable
called stratum in the expanded data set for this purpose). The expanded data set
created by stcompadj can be saved and read into Stata for us to use for further
analyses. Run the code below to generate and store an expanded data set, use it to
fit a Cox regression with the relevant interaction. Is there statistical evidence that
the effect of sex is different for the two outcomes?
preserve
stcompadj sex=1 , compet(2) savexp(silong,replace)
use silong,clear
xi:stcox i.sex*i.stratum, strata(stratum) nohr nolog
restore
(q) Lastly, we will now refit a Cox model where the effect of sex in no longer assumed
to be shared between the outcomes. This can be done by reading in the expanded
data set and fitting a Cox model directly to it. Based on the model output, what
is the estimated hazard ratio for sex on cancer-specific and other-cause mortality,
respectively?
preserve
stcompadj sex=1 , compet(2) maineffect(sex) competeffect(sex) savexp(silong,replace)
use silong,clear
xi: stcox Main_sex Compet_sex stratum, nolog
restore
55
150. Adjusted/standardized survival curves
Stata addon required! This exercise requires the Stata user-written command stpm2
This question uses the Rotterdam breast cancer data, comprising information on 2,982
patients with primary breast cancer. The outcome will be relapse-free survival, which
is defined as the time from primary surgery to disease recurrence or death from breast
cancer. Time to death from other causes were treated as censored.
(a) Load and stset the data. Restrict the follow-up time to 10 years.
. use rott2
. stset rf, f(rfi==1) scale(12) exit(time 120)
Plot the Kaplan-Meier estimate of the survival function by hormonal treatment group
(no hormonal therapy vs hormonal therapy).
. sts graph, by(hormon)
. sts gen S_km = s, by(hormon)
Will the unadjusted hazard ratio for hormonal therapy be less than or greater than
1?
(b) Now fit a proportional hazards flexible parametric model using stpm2. Use 3 df for
the baseline.
. stpm2 hormon, scale(hazard) df(3) eform
(c) Compare the model-based and Kaplan-Meier survival curves. Comment on the extent
of agreement between the two (did you expect agreement? is there agreement).
. twoway (line S_km _t if hormon == 0, sort lcolor(black) lpattern(dash) ///
connect(stepstair)) ///
(line S_km _t if hormon == 1, sort lcolor(red) lpattern(dash) ///
connect(stepstair)) ///
(line s _t if hormon==0,sort lcolor(black) lwidth(thick)) ///
(line s _t if hormon==1, sort lcolor(red) lwidth(thick)) ///
, xtitle("Years from surgery") ///
ytitle("S(t)") ///
legend(order(3 "No hormonal therapy" 4 "hormonal therapy") ring(0) ///
pos(1) cols(1)) caption("Dashed lines show KM estimates")
(d) In a previous analysis of this data, it was proposed to incorporate the effect of the
number of positive lymph nodes using the following transformation[9].
Add enodes in the model.
. stpm2 hormon enodes, scale(hazard) df(3) eform
What has happened to the hazard ratio for hormon?

(e) Now add further covariates to the model. Include the effect of age (as a restricted
cubic spline with 3 df), and tumour size.
. rcsgen age, df(3) gen(agercs) orthog
. stpm2 hormon i.size enodes agercs*, scale(hazard) df(3) eform
56 EXERCISES
(f) Obtain the predicted survival function at 1 year and 5 years. Produce a histogram
for each measure.
. gen t1 = 1
. gen t5 = 5
. predict s1, surv timevar(t1)
. predict s5, surv timevar(t5)
. hist s1, name(hist_1yr, replace) xlabel(0(0.1)1)
. hist s5, name(hist_5yr, replace)xlabel(0(0.1)1)
(g) Predict a prognostic index. This is the predicted values of the linear predictor without
the spline terms. This can be used to classify into risk groups. We will plot from
the 10th to the 90th centile of the prognostic index to show the range in predicted
survival probability in the study population.
First predict the prognostic index and then refit the model with this as the only
covariate.
. predict xb, xbnobaseline
. stpm2 xb, scale(h) df(3)
Compare the likelihood to the previous model. Why are they the same?
Now obtain predictions from the 10th to the 90th centile and plot the resulting
functions.
. forvalues i = 10(10)90 {
centile xb, centile(i)
predict s_xbi, surv at(xb r(c_1))
}
. twoway (line s_xb?? _t, sort lcolor(black ..)) ///
, legend(off) ///
ylabel(0(0.2)1, angle(h)) ///
xtitle("Years from surgery") ///
ytitle("S(t)") ///
text(0.8 8 "10th centile") ///
text(0.1 8 "90th centile")
(h) We will now move on to obtaining adjusted survival curves. Refit the original model
and and obtain the average survival curve for the study population as a whole. Note
that in large datasets the prediction can take a long time so use the timevar option
to ensure that predictions are made at selected values of time.
. stpm2 hormon i.size enodes agercs*, scale(hazard) df(3) eform
. range timevar 0 10 100
. predict s_mean, meansurv timevar(timevar) ci
. twoway (rarea s_mean_lci s_mean_uci timevar, sort pstyle(ci)) ///
(line s_mean timevar, sort) ///
ytitle("S(t)") ///
legend(off)
57
(i) Obtain the adjusted survival curves by hormonal therapy status standardising over
the covariate pattern of the whole study population. Use the meansurv option com-
bined with the at() option.
. predict s_h0, meansurv at(hormon 0) timevar(timevar) ci
. predict s_h1, meansurv at(hormon 1) timevar(timevar) ci
. twoway (line s_h0 timevar, sort) ///
(line s_h1 timevar, sort) ///
ytitle("S(t)") ///
ylabel(0(.2)1,angle(h)) ///
pos(1) cols(1)) name(adj1, replace)
(j) Obtain the adjusted survival curves by hormonal therapy status standardising over
the covariate pattern of those not on hormonal therapy.
. predict s_h0b if hormon==0, meansurv at(hormon 0) timevar(timevar) ci
. predict s_h1b if hormon==0, meansurv at(hormon 1) timevar(timevar) ci
. twoway (line s_h0b timevar, sort) ///
(line s_h1b timevar, sort) ///
ytitle("S(t)") ///
(k) Obtain the adjusted survival curves by hormonal therapy status standardising over
the covariate pattern of those on hormonal therapy.
. predict s_h0c if hormon==1, meansurv at(hormon 0) timevar(timevar) ci
. predict s_h1c if hormon==1, meansurv at(hormon 1) timevar(timevar) ci
. twoway (line s_h0c timevar, sort) ///
(line s_h1c timevar, sort) ///
ytitle("S(t)") ///
Explain why the curve in parts (j) and (k) look so different?
(l) Now calculate and plot the difference in adjusted survival curves. To do this you can
use the predictnl command. This is a very powerful postestimation Stata command
that performs the delta method in order to obtain standard errors and 95% confidence
intervals of complex functions of model parameters.
. predictnl sdiff = predict(meansurv at(hormon 0) timevar(timevar)) - ///
predict(meansurv at(hormon 1) timevar(timevar)) ///
, ci(sdiff_lci sdiff_uci)
twoway (rarea sdiff_lci sdiff_uci timevar, sort pstyle(ci)) ///

(line sdiff timevar, sort) ///
ytitle("Difference in S(t)") ///
legend(off)
58 EXERCISES
180. Outcome-selective sampling designs (nested case-control and case-cohort)
In this exercise we compare a full cohort analysis of the melanoma data to analyses using
nested case-control (NCC) and case-cohort designs. For the purpose of the exercise, we
will assume that the main exposure of interest is sex and we will adjust for age at diagnosis
(agegrp), year of diagnosis (year8594) and stage.
We would not use outcome-selective sampling in practice for this research question, but
do it here for pedagogic purposes. In practice, we might use such designs if we were
interested in collecting additional information on an expensive or time-consuming exposure
or confounder/effect modifier (e.g., biomarker information or collecting information from
medical records), which may not be feasible on the full cohort of 7,775 patients.
(a) We start by performing a full cohort analysis on all 7,775 patients. We stset using
death due to melanoma as the outcome and time-since-diagnosis as the timescale.
. stset exit, fail(status==1) enter(dx) origin(dx) scale(365.24) id(id)
How many deaths are there among the patients? These deaths will be the cases in
the NCC and the case-cohort designs.
(b) Is there evidence of a difference in mortality between women and men? Estimate
Kaplan-Meier curves and fit a Cox regression model with sex as the main exposure.
Also adjust the model for age, year and stage. Is there evidence of confounding by
age, year and stage?
. * Kaplan-Meier curves
. * Cox regression
. stcox i.sex
. stcox i.sex i.agegrp i.year8594 i.stage
(c) Now fit the same model, but as a flexible parametric model and as a Poisson regression
model. To adjust for time-since-cancer as the underlying timescale in the Poisson
regression model, we must time-split the data on follow-up (fuband) and add time-
since-cancer in the model as a covariate (fuband). This step is not needed for the
flexible parametric model, which automatically uses the timescale specified in stset.
* Flexible parametric model
. stpm2 i.sex i.agegrp i.year8594 i.stage, df(5) scale(hazard) eform
* Poisson regression
. stsplit fuband, at(0(5)20)
. streg i.sex i.agegrp i.year8594 i.stage i.fuband, dist(exp)
Compare the estimates of the sex effect for Cox regression, FPM and Poisson regres-
sion. Are they different? Would you expect them to differ? Why (or why not)?
59
You may wish to make use of Statas estimate machinery for storing, manipulating,
and displaying estimation results.
use melanoma, clear
stset exit, fail(status==1) enter(dx) origin(dx) scale(365.24) id(id)
stcox i.sex i.agegrp i.year8594 i.stage, nolog

estimates store cox
stpm2 i.sex i.agegrp i.year8594 i.stage, df(5) scale(hazard) eform nolog

estimates store fpm
stsplit fuband, at(0(5)20)

streg i.sex i.agegrp i.year8594 i.stage i.fuband, dist(exp) nolog
estimates store poiss
estimates table cox fpm poiss, eq(1) b(%5.3f) eform
(d) We will now generate and analyse a nested case-control study with 1 control per case.
First reload and stset the data using melanoma-specific death as the outcome and
time-since-diagnosis as the timescale.
How many events (deaths due to melanoma) were observed during follow-up. If we
generated a nested case-control study with 1 control per case, how many unique
individuals do you expect would be in the NCC?
Lets now generate the NCC using the sttocc command. We will match on age group
and select 1 control per case.
. set seed 339487731
. sttocc, match(agegrp) n(1)
The sampling will take a few minutes. For each riskset a dot will appear in the results
window. Since there are 1913 deaths, there will be 1913 dots. We have chosen to
specify a seed for the Stata random number functions in order to make the sampling
reproducible (i.e., force everyone to get the same results).
(e) The data set in memory will now be the nested case-control dataset. Use the describe
command to examine its contents and list the first 10 observations.
The variable case include the case-control status (1=case, 0=control), set is a
unique identifier for the matched set (i.e., the case and its matched control will have
the same value on the set variable).
i. What information is represented by the variable time?
ii. Confirm that there are an equal number of cases and controls and that the age
distribution of the cases and controls is the same (due to matching on age).
iii. How many unique individuals are there in the nested case-control study?
60 EXERCISES
(f) Nested case-control studies are analysed using conditional logistic regression. We
must condition on the matching strata (by including the set variable in the group()
option).
. clogit _case i.sex i.year8594 i.stage, group(_set) or
i. What underlying quantity is being estimated by the estimates in the column

labelled Odds ratio ?
ii. Is the estimated effect measure for sex similar to the hazard ratio for the full
cohort? Would you expect it to be?
iii. Are the confidence intervals (standard errors) similar?
(g) We will now generate and analyse a case-cohort study. Reload and stset the data
using cause-specific death as the outcome and time-since-diagnosis as the timescale

Now we will generate a case-cohort design with a sampling fraction of around 25%.
We will generate a new outcome variable based on the event indicator variable ( d)
from the stset.
. gen case=_d
First, we must sample the subcohort. To make sure the sampling is reproducible, we
use a seed. Then we assign a random number between 0 and 1 to all observations
in the dataset (using the runiform command). We select a subcohort by creating
an indicator variable subcoh which takes the value 1 for 25% observations in the
subcohort and value 0 for the remaining 75% observations outside the subcohort.
. set seed 339487731 // makes sampling reproducible

. gen u = runiform() // assign random number to all obs
. gen subcoh = 1 if (u <= 0.25) // generate dummy subcohort
. replace subcoh = 0 if (u > 0.25)
Check that the sampling worked by tabulating the number of cases and non-cases
inside and outside the subcohort? Complete the following table.
. tab case subcoh
Outside subcohort Inside subcohort Total

Non-cases
Cases
Total 7,775
(h) Calculate the exact sampling fraction of the subcohort. Also calculate the exact sam-
pling fraction of non-cases, i.e. the proportion of non-cases in the subcohort compared
to non-cases in the full cohort.
61
(i) The analysis of case-cohort samples is identical to that of cohort designs with the
addition of (1) weights and (2) robust standard errors. Generate the weights by
creating a wt variable, which takes the value 1 for cases and the inverse of the
sampling fraction for non-cases for non-cases.
. gen wt = 1 if case==1
. replace wt = 1 / (1470/5862) if case==0 & subcoh==1
. tab wt, missing
Are the weights as you expected? Which observations have missing values for wt?
(j) To include weights in the analysis in Stata, simply include them in the stset com-
mand using the pweight option [pw=wt]. The weights will now be included in all st
commands. Any observation with missing values for wt will not be included in the
analysis.
. stset exit [pw=wt], fail(status==1) enter(dx) origin(dx) ///
scale(365.24) id(id)
(k) Estimate the effect of sex on mortality by fitting the models using the weighted data,
i.e. Cox regression, FPM and Poisson regression. Compare the estimates of the sex
effect. What do you conclude?
. * Cox model for case-cohort - Borgan II weights
. stcox i.sex i.agegrp i.year8594 i.stage, vce(robust)
. * FPM model for case-cohort - Borgan II weights

. stpm2 i.sex i.agegrp i.year8594 i.stage, scale(h) df(5) eform ///
vce(robust) nolog
. * Poisson regression - Borgan II weights

. stsplit fuband, at(0(5)20)
. streg i.sex i.agegrp i.year8594 i.stage i.fuband, dist(exp) vce(robust)
(l) Investigate the extent of sampling variation by generating multiple nested case-control
studies. The following code generates, analyses, and reports a table of results for 5
repetitions. Note that this code will take sevarl minutes to run.
set more off
use melanoma, clear
stset exit, fail(status==1) enter(dx) origin(dx) scale(365.24) id(id)
forvalues i=1/5 {
preserve
display as text _newline "Now processing iteration " i _newline
sttocc, match(agegrp) n(1)
clogit _case i.sex i.year8594 i.agegrp i.stage, group(_set) or nolog
estimates store ncci
restore
}
est table ncc1 ncc2 ncc3 ncc4 ncc5, eform equations(1) ///
b(%9.6f) se modelwidth(10)
(m) Now see what happens if you increase the number of controls to, for example, 3 and
then 5. What would you expect with 5 controls per case?
62 EXERCISES
(n) Investigate generate and analyse multiple case-cohort studies. The following code
generates, analyses, and reports a table of results for 5 repetitions with a subcohort
of 25%.
set more off
use melanoma, clear
gen case=(status==1)
forvalues i=1/5 {
preserve
display as text _newline "Now processing iteration " i _newline
gen subcoh = (runiform() <= 0.25)
gen wt = 1 if case==1
replace wt = 1 / 0.25 if case==0 & subcoh==1
stset exit [pw=wt], fail(status==1) enter(dx) origin(dx) scale(365.24) id(id)
stcox i.sex i.agegrp i.year8594 i.stage, vce(robust)
estimates store cci
restore
}
est table full_cox cc1 cc2 cc3 cc4 cc5, eform equations(1) ///
b(%9.6f) se modelwidth(10)
(o) Now investigate the effect of changing the size of the subcohort to, for example, 10%
and 50%.
63
181. Calculating SMRs/SIRs
The standardized mortality ratio (SMR) is the ratio of the observed number of deaths
in the study population to the number that would be expected if the study population
experienced the same mortality as the standard population. It is an indirectly standardized
rate. When studying disease incidence the corresponding quantity is called a standardized
incidence ratio (SIR). These measures are typically used when the entire study population
is considered exposed. Rather than following-up both the exposed study population and
an unexposed control population and comparing the two estimated rates we instead only
estimate the rate (or number of events) in the study population and compare this to the
expected rate (expected number of events) for the standard population. For example, we
might study disease incidence or mortality among individuals with a certain occupation
(farmers, painters, airline cabin crew) or cancer incidence in a cohort exposed to ionising
radiation.
In the analysis of cancer patient survival we typically estimate excess mortality (observed -
expected deaths). The SMR (observed/expected deaths) is a measure of relative mortality.
The estimation of observed and expected numbers of deaths are performed in an identical
manner for each measure but with the SMR we assume that the effect of exposure is
multiplicative to the baseline rate whereas with excess mortality we assume it is additive.
Which measure, relative mortality or excess mortality, do you think is more homogeneous
across age?
The following example illustrates the approach to estimating SMRs/SIRs using Stata.
Specifically, we will estimate SMRs for the melanoma data using the general population
mortality rates stratified by age and calendar period (derived from popmort.dta) to es-
timate the expected number of deaths. The expected mortality rates depend on current
age and current year so the approach is as follows
Split follow-up into 1-year age bands

Split the resulting data into 1-year calendar period bands
For each age-period band, merge with popmort.dta to obtain the expected mortality
rates
Sum the observed and expected numbers of deaths and calculate the SMR (ob-
served/expected) and a 95% CI
(a) Start by stsetting the data with age as the timescale and splitting the follow-up into
1 year age bands
. stset exit, fail(status == 1 2) origin(bdate) entry(dx) scale(365.24) id(id)
. stsplit _age, at(0(1)110) trim
(b) Now split these new records into 1 year calendar period bands using
. stsplit _year, after(time=d(1/1/1975)) at(0(1)22) trim
. replace _year=1975+_year
. list id _age _year in 1/15
Note that we have used the second syntax for stsplit and set the origin for calendar
period as 1/1/1975 for convenience in setting the breaks.
64 EXERCISES
(c) Each subjects followup is now divided into small pieces corresponding to the age-
bands and calendar periods the subject passes through. We can make tables of deaths
and person-years by age and calendar period with
. gen _y = _t - _t0 if _st==1
. table _age _year, c(sum _d)
. table _age _year, c(sum _y) format(%5.3f)
As the data have been split in 1-year intervals on both time scales the table created
above is not so informative. Grouped variables will provide a better overview.
. egen ageband_10=cut(_age), at (0(10)110)
. egen period_5=cut(_year), at(1970(5)2000)
. table ageband_10 period_5, c(sum _d)

. table ageband_10 period_5, c(sum _y) format(%4.1f)
(d) To make a table of rates by age and calendar period, try
. gen obsrate=_d/_y
. table ageband_10 period_5 [iw=_y] , c(mean obsrate) format(%5.3f)
(e) To calculate the expected cases for a cohort, using reference mortality rates classified
by age and calendar period, it is first necessary to break the followup into parts
which correspond to these age bands and calendar periods, as above.
Before calculating the expected number of cases it is necessary to add the reference
rates to the expanded data with
. sort _year sex _age
. merge m:1 _year sex _age using popmort
This is a matched merge on age band and calendar period and will add the appropriate
survival probability to each record. The system variable merge takes the following
values:
1 record in the master file but no match in popmort
2 record in popmort but no match in the master file
3 record in the master file with a match in popmort
. tab _merge
should show mostly 3s with some 2s but no 1s. You can now drop the records with
no match in the master file and the system variable
. drop if _merge==2
. drop _merge
(f) The mortality rates for the standard population are derived by transforming the
survival probabilities
. gen mortrate=(-ln(prob))
and to calculate the expected number of cases, multiply the follow-up time for each
record by the reference rate for that record
. gen e=_y*mortrate
. list id e _d in 1/15
65
(g) The SMR is the ratio of the total observed cases to the total number expected. The
total numbers are obtained through
. egen obs=total(_d)
. egen exp=total(e)
from which the manually calculated SMR with corresponding 95% confidence inter-
val [10] are obtained (preserve and restore are used to speed up the processing)
. preserve
. keep in 1
. gen SMR = obs/exp
. gen LL = ( 0.5*invchi2(2*obs, 0.025)) / exp
. gen UL = ( 0.5*invchi2(2*(obs+1), 0.975)) / exp
. display "SMR(95%CI)=" round(SMR,.001) " (" round(LL,.001) ":" round(UL,.001) ")"

. restore
An easier approach is to let the strate command perform these calculations for us
(after first splitting and merging in the standard rates).
. strate, smr(mortrate)
(h) To calculate the SMR for the different stages, try
. strate stage, smr(mortrate)
Summary
The following commands can be used to calculate an SMR for the melanoma patients:

. merge _year sex _age using popmort
. drop if _merge==2
. gen mortrate=-ln(prob)
66 EXERCISES
182. Using strs for calculating SMRs
Stata addon required! This exercise requires the Stata user-written command strs.
The Stata command for estimating relative survival (strs) can also be used for calculating
SMR. The primary purpose of this command is to estimate excess mortality (observed
minus expected deaths). We can therefore use strs to estimate the numbers of observed
and expected deaths for us and then calculate the ratio of these two numbers to get an
SMR/SIR. That is, strs does the splitting and merging for us. We start by stsetting the
data with time since diagnosis as the time scale.

. stset exit, fail(status == 1 2) origin(dx) entry(dx) scale(365.24) id(id)
Note that strs expects there to be a variable in the dataset containing age at diagnosis
and calendar year at diagnosis (the default names for these variable are age and yydx)
and uses these variables to keep track of current age and current year.
. strs using popmort, br(0(1)21) mergeby(_year sex _age) notables save(replace)
strs saves two datasets: one containing individual subject-band data (individ.dta) and
one with grouped life-table data (grouped.dta). Either can be used for calculating the
SMR and confidence interval but using the grouped data is faster.
We suppressed the printing of the lifetable in the previous step (using the notables option)
but the same information is stored in grouped.dta.
. use grouped.dta, clear

. list start n d w p cp d_star, sum(d d_star)
The observed number of events is in the variable d and the expected number of events is
in the variable d_star. Now we can sum these variables and take the ratio to obtain the
SMR.
. collapse (sum) obs=d exp=d_star

. gen LL=( 0.5*invchi2(2*obs, 0.025)) / exp
. gen UL=( 0.5*invchi2(2*(obs+1), 0.975)) / exp
. gen smr=obs/exp
. list obs exp smr LL UL
The results obtained using strs are not identical to those obtained in question 181. This
is because in question 181 we split on both current age and current year whereas strs
only splits on age and assumes that current year increments by one each and every time
current age increments by one. Using the calyear option with strs will cause it to also
split on calendar year, thereby giving the exact same results as in question 181. Splitting
on calendar year increases the computational time and is therefore not done by default.
67
The following code illustrates the differences.

Splitting on both current age and current year.
. gen bdate = dx-(age*365.25)


. drop if _merge==2
Splitting only on current age and approximating current year.
. gen bdate = dx-(age*365.25)
. gen _year=year(dx)+(_age-age)

. drop if _merge==2
If for example, a man is diagnosed with cancer in June 1990 aged 81 then the approximate
approach assumes that for the entire first year of follow-up he will experience the 1990 rates
and for the entire second year of follow-up he will experience the 1991 rates. The exact
approach (splitting on both age and period) results in this man, during the second year of
follow-up experiencing 1991 rates for 6 months and 1992 rates for 6 months. If population
mortality is decreasing over time then the approximation will result in an overestimate of
the expected number of deaths. This bias should not, however, be of practical significance.
+----------------------------------------+
| D E SMR Lower Upper |
|----------------------------------------|
Exact approach | 3047 1261.02 2.416 2.332 2.504 |
Approximation | 3047 1268.08 2.403 2.319 2.490 |
+----------------------------------------+
68 EXERCISES
200. Calculating expected survival by hand (with help from MS Excel)

The Microsoft Excel file expected.xls contains annual probabilities of survival for mem-
bers of the general population matched to the sample of 35 patients diagnosed with colon
carcinoma.
These annual probabilities of survival were obtained by matching with the data in popmort.dta
(a Stata data file).
(a) Five expected probabilities are listed for patient number 1. Locate the corresponding
probabilities in popmort.dta.
(b) The expected probabilities are missing for patient number 4. Complete the missing
cells using the data in popmort.dta.
(c) Calculate (using Excel) the five-year expected survival probabilities for the 35 patients
using both the Ederer I and Ederer II methods.
(d) Confirm your results using strs.
. use colon_sample, clear
. gen id = _n
. stset surv_mm, failure(status==1 2) scale(12) id(id)
. strs using popmort, breaks(0(1)5) mergeby(_year sex _age) ///
ederer1 list(n d w p cp_e1 cp_e2)
69
201. Localised melanoma: life table estimates of relative survival by calendar period
of diagnosis using strs
Use strs to calculate life table estimates of relative survival (Ederer II) for patients diag-
nosed with localised skin melanoma for each of the two calendar periods 19751984 and
19851994. Use annual intervals up to 10 years. Following is the required Stata code. It
is strongly suggested that you run these commands from a do file.

. stset surv_mm, fail(status==1 2) id(id) scale(12)
. strs using popmort, br(0(1)10) mergeby(_year sex _age) by(year8594) save(replace)
(a) Confirm that you understand what each of columns represents and check, for exam-
ple, that the relative survival is equal to the ratio of observed to expected survival.
Confirm that cumulative relative survival is equal to the product of interval-specific
survival (for the first couple of intervals).
i. During which point in the follow-up is excess mortality highest?
ii. Does the peak in excess mortality occur where you would expect from a biologi-
cal/clinical perspective?
iii. Is there evidence that the patients reach a point of statistical cure?
(b) Now estimate the life table using intervals of length 6 months rather than 1 year.
Does the estimate of 10-year relative survival change a great deal as a result?
(c) Now estimate the life table using intervals of length 3 months for the first year followed
by intervals of 1 year. Does the estimate of 10-year relative survival change a great
deal as a result?
(d) Now estimate 20-year relative survival (using annual intervals). Why is it not possible
to estimate 20-year survival for both periods?
(e) Plot the estimates of cumulative relative survival for each calendar period against
follow-up time (estimate survival using annual intervals up to at least 20 years).
. use grouped, clear
. twoway (connected cr end if year8594==0) (connected cr end if year8594==1)
Following is some alternative code for producing a graph (for you to copy and paste
from the PDF file into the do file editor).
. twoway (scatter cr end if year8594==0, msymbol(O)) ///
(scatter cr end if year8594==1, msymbol(O)) ///
(rcap lo_cr hi_cr end if year8594==0, lcolor(black)) ///
(rcap lo_cr hi_cr end if year8594==1, lcolor(black)) , ///
yti("Relative Survival") yscale(range(0.4 1)) ///
ylabel(0.4(0.2)1, format(%3.1f)) ///
xti("Years from diagnosis") xla(0(2)20) ///
legend(order(1 "1975-84" 2 "1985-94") ring(0) pos(7) col(1))
70 EXERCISES
(f) Plot the estimates of interval-specific relative survival for each calendar period against
follow-up time.
. twoway (rcap lo_r hi_r end) (scatter r end), ///
by(year8594, legend(off)) yti("Interval-specific RSR") ///
xti("Years from diagnosis") xla(0(2)20) yla(0.8(0.1)1)
Revisit questions (i)(iii) in part (a). Do you gain additional insight if you redraw
the graph with narrower (e.g., 6 month) intervals? Note that to change the interval
widths you will need to go back to the patient data and rerun strs.
(g) Estimate and compare cumulative relative survival (stratified by calendar period)
using three alternative methods for estimating expected survival (Hakulinen, Ederer I,
and Ederer II).
. stset exit, origin(dx) fail(status==1 2) id(id) scale(365.24)
. gen long potfu = date("31/12/1995","DMY") /* "dmy" if using Stata 9 */
. strs using popmort if stage==1, br(0(1)20) mergeby(_year sex _age) ///
by(year8594) list(start n d w cr_e1 cr_e2 cr_hak) ederer1 potfu(potfu)
Are there large differences in the estimates of cumulative relative survival depending
on the method used to estimate expected survival (Hakulinen, Ederer I, and Ed-
erer II)? Study the differences at 1, 5, 10, 15, and 20 years subsequent to diagnosis.
(h) Use the pohar option to get estimates of net survival using the Pohar Perme et al. [11]
approach and compare the estimates to relative survival (Ederer II).
. stset exit, origin(dx) fail(status==1 2) id(id) scale(365.24)
. strs using popmort, br(0(=1/12)20) mergeby(_year sex _age) ///
by(year8594) pohar list(start n d w cr_e2 cns_pp) save(replace)
. use grouped, clear
. list start end cr_e2 cns_pp if mod(end,1)==0 & year8594, noobs
(i) Use strs to generate a single lifetable for all patients diagnosed with localised skin
melanoma (annual intervals up to 10 years). Use the save option, that will cause the
results to be saved to grouped.dta (one observation for each life table interval) and
individ.dta (one observation for each person for each life table interval).
i. Use the data saved to grouped.dta to estimate observed, expected, and excess
mortality rates for each life table interval. Before doing the calculations, consider
how you expect expected mortality to vary as a function of time since diagnosis
(i.e., do you expect it to increase, decrease, or remain constant). Was the pattern
what you expected? We will investigate this further in part (iii) by looking at
the individual data.
ii. Plot the excess mortality rates as a function of time since diagnosis. Is the pattern
similar to the pattern for cause-specific mortality (e.g., the graph you plotted in
exercise 111f)?
iii. Open the individual data and calculate the average age of the patients during
each interval.
. use individ, clear
. collapse (mean) age _age, by(end)
. list
_age is the average age of the patients still at risk at the start of each interval
and age is the average age of these same patients were at diagnosis. Here we
see how the age distribution of the patients varies with follow-up and that the
elderly are more likely to die earlier.
71
202. Localised melanoma: life table estimates of cause-specific survival by calendar

period of diagnosis using ltable and strs
The purpose of this exercise is to illustrate how the strs command can be used for es-
timating cause-specific survival (results should be identical to those obtained from the
ltable command) and to compare estimates of cause-specific to relative survival.
(a) Use strs to estimate cause-specific survival (using annual life table intervals) for
patients diagnosed with localised skin melanoma (a single life table for all patients).
HINT: To estimate cause-specific survival using strs you will need to define the
failure event as death due to melanoma when you stset the data. The figures in
the CP column will then be the cumulative cause specific survival. The estimates of
relative survival (the R and CR columns) will be meaningless.
(b) Repeat the previous exercise but instead use the ltable command. Confirm that
both commands give identical estimates.
HINT: Use the online help if you are unsure of the command syntax. You will need
to create an indicator variable for cause-specific death. Note that ltable is not an
st command (so does not require or respect stset).
(c) Compare the estimated relative survival ratios to the estimated cause-specific survival
rates. Would you expect substantial differences? What could explain the differences?
72 EXERCISES
203. Localised melanoma: period estimation of relative survival
The commands in q203.do produce period estimates (using the window 01jan199431dec1995)
of relative survival for each sex. Check that you understand each of the commands in the
do file and compare the period estimates of survival with cohort estimates of relative
survival.
73
204. Localised melanoma: period estimation of relative survival using data up to

the end of 1983
Imagine it is early 1984. You have data for patients diagnosed and followed up until the
end of 1983 and wish to predict the 5-year relative survival of newly diagnosed patients.
We will calculate period estimates and cohort estimates and compare these to the actual
survival of patients diagnosed in 1984.
(a) Using the knowledge that patient survival is improving with time:
i. which do you expect will be higher; the period estimate or the cohort estimate
of relative survival?
ii. which do you expect will be a better predictor of the survival of patients diagnosed
in 1983; the period estimate or the cohort estimate?
(b) Calculate the 5-year relative survival (Ederer II, annual life table intervals) using the
traditional cohort method based on patients diagnosed 1975-1983 followed up until
the end of 1983. Throughout this exercise you should specify the dates of diagnosis
and exit when using stset.
(c) Calculate the 5-year relative survival (Ederer II, annual life table intervals) using
the traditional cohort method based on patients diagnosed 1977-1983 followed up
until the end of 1983. That is, repeat the last part but without including patients
diagnosed 1975 & 1976. Before performing the calculations, use your knowledge of
how survival is changing over time to predict whether the estimated survival will be
lower or higher than the estimate you obtained in the previous part.
(d) Calculate the 5-year relative survival (Ederer II, annual life table intervals) using
period analysis with a window of 01jan198331dec1983.
(e) Calculate the 5-year relative survival (Ederer II, annual life table intervals) using
period analysis with a window of 01jan198231dec1983. How would you expect the
estimates to differ from the previous part?
(f) Calculate the actual 5-year relative survival for patients diagnosed 1983.
(g) Calculate the actual 5-year relative survival for patients diagnosed 1984.
(h) Compare the estimates of 5-year relative survival (and associated standard errors)
calculated in the previous parts. Did the relative magnitudes of the estimates conform
to your expectations?
(i) If relative survival does not vary according to calendar period of diagnosis, which of
the following is most likely to be true?
i. the period estimate of relative survival will be lower than the cohort estimate.
ii. the period estimate of relative survival will be equal to the cohort estimate.
iii. the period estimate of relative survival will be higher than the cohort estimate.
(j) ADVANCED: For each and every calendar year between 1981 and 1990, calculate the
following and plot them on a graph (with calendar year on the X axis)
i. the most up-to-date cohort estimate of 5-year relative survival for patients diag-
nosed up until the end of the year.
ii. the most up-to-date period estimate of 5-year relative survival for patients diag-
nosed up until the end of the year.
iii. the actual 5-year relative survival for patients diagnosed during the subsequent
year.
74 EXERCISES
210. Model excess mortality using Poisson regression
We will now model relative survival (excess mortality) in patients diagnosed with localised
melanoma as a function of follow-up time, sex, age at diagnosis, and period for the first 5
years of follow-up. The first step is to estimate relative survival for each combination of
sex, age at diagnosis, and period which can be done by running the commands in q210.do.
Be sure to study this file in the Stata do-file editor (or some other text editor) to ensure
you understand the commands it contains.
We can model excess mortality using Poisson regression [12] using the following commands.
. use grouped if end < 6, clear

. glm d i.end i.sex i.year8594 i.agegrp, fam(pois) link(rs d_star) lnoff(y) eform
The eform option requests that the estimates be presented as the exponential of the
estimated parameters (i.e. relative excess risks), rather than the estimated parameters.
(a) During which year of follow-up was excess mortality highest? Is this what you would
expect?
(b) Compare the estimates (and their interpretation) to those obtained from the Cox
model for cause-specific mortality (and also to the Poisson regression model for cause-
specific mortality if you did it).
(c) Fit the model without the eform option and ensure you understand why some values
in the output change and some do not. Note that you can obtain these estimates
simply by typing
. glm
which results in Stata displaying the estimates from the last model estimated using
the glm command.
(d) Test the assumption of proportional excess hazards across age groups by fitting an
appropriate interaction term in the model.
(e) The model in the previous part was estimated based on collapsed data (d, d_star and
y were summed across covariate patterns). Now fit the model to individual subject-
band data by using the data set individ rather than grouped. Do the estimates
change considerably? Would you expect them to?
(f) Now model relative survival using the full likelihood approach (Estève et al. [13]) and
compare the estimates with the Poisson regression model estimated using individual
data (the estimates should be identical).
. use individ if end < 6, clear
. ml model lf esteve (d=i.end sex year8594 i.agegrp)
. ml maximize, eform("RER")
(g) Now model relative survival using the Hakulinen-Tenkanen approach [14] and compare
the estimates with previous models.
. use grouped if end < 6, clear
. glm ns i.end i.sex i.year8594 i.agegrp, fam(bin n_prime) link(ht p_star) eform
(h) Time permitting, fit similar models to the data for all stages (include stage in the
model). Test the assumption of proportional excess hazards for stage.
75
211. Model excess mortality using Poisson regression with a smooth baseline
Stata addons required! This exercise requires the Stata user-written command rcsgen.
Type ssc install rcsgen to install.
We will now extend the Poisson modelling approach by defining narrow intervals and fitting
a smooth function (e.g., splines or fractional polynomials) of time for the excess hazard
rate rather than a piecewise estimate [15, 16].
It is possible to use individual level data in these situations, but having many rows of data
per subject can lead to very large file sizes and computation can be very slow (less so now
than when this exercise was first written). If you are using categorical covariates then you
can collapse the data over these covariates and follow-up interval. For large datasets this
is particularly useful.
(a) Load the melanoma data and use strs to split the time scale using 1 month inter-
vals. Use the option by(agegrp sex year8594) as these are the covariates we are
interested in modelling.
. use melanoma
. stset surv_mm, failure(status=1 2) scale(12) exit(time 120.5) id(id)
. strs using popmort, br(0(0.08333)6) mergeby(_year sex _age) ///
by(agegrp sex year8594) notables ///
savind(vnarrowint_ind, replace) ///
savgroup(vnarrowint_grp, replace)
Open the vnarrowint ind and vnarrowint grp data file and compare the number
of observations in each dataset.
For computational speed we will use the grouped data.
(b) Create new variables for the average follow-up time for each interval, dummy variables
for the covariates of interest and the restricted cubic spline variables (for the moment
we will have knots at 0.05, 0.25, 0.75, 1.5, 2.5, and 4 years.
. use vnarrowint_grp, clear
. gen midtime = (start+end)/2

. rcsgen midtime, knots(0.05 0.25 0.75 1.5 2.5 4) gen(rcs)
Fit a proportional excess hazards model using restricted cubic splines for the baseline
hazard.
. glm d rcs1-rcs5 agegrp2 agegrp3 agegrp4 female year8594 ///
, family(poisson) link(rs d_star) lnoffset(y)
. estimates store M_sp_peh
Use the glm, eform option to obtain excess hazard ratios. Compare these to the
estimated excess hazard ratios obtained in exercise 210 where follow-up time was
modelled using a step function.
76 EXERCISES
(c) Calculate predicted values for the excess hazard rate and plot for the oldest and
youngest groups (for males in the 1985-1994 period of diagnosis).
. predict lh, xb nooffset
. gen h=exp(lh)
. twoway (line h midtime if agegrp1 == 1 & female == 0 & year8594 == 1, sort) ///
(line h midtime if agegrp4 == 1 & female == 0 & year8594 == 1, sort)
Use the yscale(log) to plot the excess hazard on the log scale. Why are these lines
parallel?
(d) Create dummy variables for the interaction between the age groups and the spline
variables and then include these in the model to allow for non-proportionality of the
excess hazards (for age group).
forvalues j = 1/5 {
gen ageircsj = agegrpi * rcsj
}
}
. glm d rcs1-rcs5 agegrp2 agegrp3 agegrp4 female year8594 ///

age2rcs1-age2rcs5 age3rcs1-age3rcs5 age4rcs1-age4rcs5 ///
,family(poisson) link(rs d_star) lnoffset(y)
. lrtest M_sp_peh
Is there evidence of non-proportionality of the excess hazard rates?

(e) Obtain estimates of the time-dependent excess hazard ratios for age group. Use the
partpred command to obtain partial predictions. Plot the excess hazard ratios
against follow-up time.
. forvalue i = 2/4 {
partpred agehri, for(agegrpi ageircs1-ageircs5) eform ///
ci(agehri_lci agehri_uci)
}
. twoway (line agehr2 midtime if agegrp2==1, sort) ///

(line agehr3 midtime if agegrp3==1, sort) ///
,legend(order(1 "45-59" 2 "60-74" 3 "75+") ring(0) pos(1) cols(1)) ///
yline(1) ytitle(Excess Hazard Ratio) xtitle(Years from Diagnosis)
(f) Now plot the time-dependent excess hazard ratio with 95% confidence intervals for
the oldest age group (compared to the youngest).
. twoway (rarea agehr4_lci agehr4_uci midtime if agegrp4==1, sort pstyle(ci)) ///
(function y = 2.9, range(0 6) lcolor(black) lpattern(dash)), ///
legend(off) yline(1) ytitle(Excess Hazard Ratio)
77
(g) THIS IS AN ADVANCED QUESTION!

Use the code in the do file 211.do (part (j)) to obtain predicted relative survival curves
with 95% confidence intervals. Compare the predicted relative survival estimates for
the youngest and oldest age groups (for males in the the 1985-1994 period) to those
obtained using life tables.
This code is fairly complex; relative survival is a function of the model parameters,
RSR=exp(-cumulative hazard), and we use the the delta method (see http://www.
stata.com/support/faqs/stat/deltam.html) to obtain standard errors. The code
could be improved by rewriting in Statas matrix language, Mata.
If you complete this exercise you will gain an insight into why we prefer stpm2. There
is a lot of code in this question, much of it advanced. If you have any questions about
the code please ask (Paul Lambert wrote the code for this part and is the best
person to ask).
(h) An alternative to using splines is fractional polynomials. The following code fits a
PEH model allowing an 3 degree fractional polynomial (FP3) model for the baseline
excess hazard. Compare the excess hazard ratios with those obtain in part (b).
Compare the fitted hazard rates with those obtained in part (c).
/* Note that this will take a few minutes to fit */
. mfp glm d midtime (agegrp2 agegrp3 agegrp4 female year8594) ///
, family(poisson) link(rs d_star) lnoffset(y) eform ///
df(4, midtime:6) alpha(-1) xorder(n)
. predict lh_fp, xb nooffset

. gen h_fp = exp(lh_fp)
(i) Fractional polynomials models can also be extended to time dependent effects. The
following code will fit FP functions for the time-dependent effects. At least a linear
effect will be incorporated for the time-dependent effect, but backward selection can
be incorporated if desired using the select option.
gen ageimidtime = midtime*agegrpi
}
. mfp, df(4, midtime:6) alpha(-1) ///

xorder(n): glm d midtime ///
(agegrp2 agegrp3 agegrp4 female year8594) ///
, family(poisson) link(rs d_star) lnoffset(y) eform
(j) Using individual data takes a longer time, but the model estimates should be very
similar. If your computer is fast enough, run the code in the solution Do file to
fit a proportional excess hazards model. Compare the excess hazard ratios to those
obtained in part (b).
78 EXERCISES
230. Modelling excess mortality using flexible parametric models
We fill now fit some models with the linear predictor on the log cumulative excess hazard
scale, i.e. flexible parametric relative survival models (an extension of Royston-Parmar
survival models).
(a) Load the Melanoma data. We need to merge in the expected hazard rate at the time
of death. This can be done making use of stset.
. stset surv_mm, failure(status=1 2) scale(12) exit(time 120.5) id(id)
. gen _age = min(int(age + _t),99)
. gen _year = int(yydx + _t)
. merge m:1 _year sex _age using popmort, keep(match master)
Use stpm2 to fit a flexible parametric relative survival model (an extension of Royston-
Parmar models) incorporating the expected mortality rate. We will start by ignoring
the effects of covariates and use 3 degrees of freedom (2 internal knots). Use predict
to obtain the predicted hazard and survival functions.
. stpm2, df(3) scale(hazard) bhazard(rate)
. predict h1, hazard per (1000) ci
. predict s1, survival ci
Use the data editor to look at the data. Notice that six new variables have been cre-
ated. These are the spline variables, rcs1- rcs3 and their derivatives, d rcs1- d rcs3.
(b) Plot the predicted hazard and survival functions.
. twoway (rarea h1_lci h1_uci _t, sort pstyle(ci)) ///
(line h1 _t, sort), name(h1)
. twoway (rarea s1_lci s1_uci _t, sort pstyle(ci)) ///
(line s1 _t, sort), name(s1)
(c) Now compare the estimated hazard and survival functions using 2 4 and 6 degrees of
freedom.
. foreach df in 2 4 6 {
stpm2, bhazard(rate) df(df) scale(hazard)
predict h_dfdf, hazard ci
replace h_dfdf = h_dfdf * 1000
predict s_dfdf, survival ci
estimates store dfdf
}
. twoway (line h_df2 h_df4 h_df6 _t, sort lcolor(red blue black))
. twoway (line s_df2 s_df4 s_df6 _t, sort lcolor(red blue black))
79
Compare the models using the AIC and BIC. See exercise 131 for notes on AIC and
BIC.
. estimates stats df2 df4 df6
Which is the best fitting model?

(d) Fit a proportional excess hazards model (using 3 degrees of freedom for the baseline)
including sex, age group and period of diagnosis.
. stpm2 agegrp2 agegrp3 agegrp4 female year8594, ///
bhazard(rate) df(3) scale(hazard) eform
. predict h2, hazard per(1000) ci
Compare the estimates hazard ratios with those from other proportional excess hazard
models (piecewise, spline and fractional polynomial models).
(e) Plot the predicted excess hazard rate for the oldest and youngest groups (for males
in the 1985-1994 period of diagnosis).
. twoway (line h2 _t if agegrp1 == 1 & female == 0 & year8594 == 1, sort) ///
(line h2 _t if agegrp4 == 1 & female == 0 & year8594 == 1, sort)
(f) Extend the model by allowing time-dependent effects for age group.
. stpm2 agegrp2 agegrp3 agegrp4 female year8594, bhazard(brate) df(3) ///
scale(hazard) tvc(agegrp2 agegrp3 agegrp4) dftvc(2)
. predict h3, hazard per(1000) ci
Plot the predicted excess hazard rate for the oldest and youngest groups (for males in
the 1985-1994 period of diagnosis). Note how these differ from those obtained from
the proportional excess hazards model.
. twoway (line h3 _t if agegrp1 == 1 & female == 0 & year8594 == 1, sort) ///
(line h3 _t if agegrp4 == 1 & female == 0 & year8594 == 1, sort)
(g) Use the hrnumerator() and hrdenominator() options to obtain the time-dependent
predicted hazard ratios for age group. Plot these for age groups 2, 3 and 4.
. predict hr2, hrnum(agegrp2 1) ci
. twoway (line hr2 _t if agegrp2 == 1 & female == 0 & year8594 == 1, sort) ///
(line hr3 _t if agegrp3 == 1 & female == 0 & year8594 == 1, sort) ///
(line hr4 _t if agegrp4 == 1 & female == 0 & year8594 == 1, sort)
Plot the hazard ratio for age group 4 with 95% confidence intervals.
. twoway (rarea hr4_lci hr4_uci _t, sort pstyle(ci)) ///
(line hr4 _t, sort), yline(1) ///
xtitle("Years from Diagnosis") ///
ytitle("Excess Mortality Rate Ratio")
80 EXERCISES
(h) Now obtain and plot the difference in estimated relative survival curves for the oldest
versus the youngest age group (for males in the 1985-1994 period of diagnosis).
. predict sdiff4, sdiff1(agegrp4 1 female 0 year8594 1) ///
sdiff2(agegrp4 0 female 0 year8594 1) ci
. twoway (rarea sdiff4_lci sdiff4_uci _t, sort pstyle(ci)) ///
(line sdiff4 _t, sort), yline(0) legend(off) ///
ytitle("Difference in Relative Survival")
(i) Now obtain and plot the difference in estimated excess mortality rates for the oldest
versus the youngest age group (for males in the 1985-1994 period of diagnosis).
. predict hdiff4, hdiff1(agegrp4 1 female 0 year8594 1) ///
hdiff2(agegrp4 0 female 0 year8594 1) ci
. replace hdiff4 = hdiff4*1000
. replace hdiff4_lci = hdiff4_lci*1000
. replace hdiff4_uci = hdiff4_uci*1000
. twoway (rarea hdiff4_lci hdiff4_uci _t, sort pstyle(ci)) ///

(line hdiff4 _t, sort), yline(0) legend(off) ///
ytitle("Difference in Excess Mortality Rate")
81
231. Modelling non-linear effects in relative survival I

Proportional hazards models
(a) Load the colon cancer data and merge in the background mortality rates as in ques-
tion 230. Only keep those aged 90 and under at diagnosis.
. keep if age<=90
(b) Fit a flexible parametric relative survival model with no covariates.
. stpm2 , scale(hazard) df(5) bhazard(rate)
It is possible to obtain Martingale-like residuals from parametric models which can
be used in a similar way to Cox models to assess the functional form of continuous co-
variates. Obtain the martingale-like residuals for this model and add a non-parametric
smoother when plotting the residuals vs age.
. predict mg1, martingale
. lowess mg1 age, name(mg1, replace)
What does this tell you about the functional form needed to model age?
(c) Now add age to the model assuming a linear effect. Interpret the coefficient for age.
(d) We can decide on a reference age and the plot the excess mortality rate ratio as a
function of age. Use the partpred command to get the predictions and confidence
intervals with age 50 as the reference.
. partpred hr_age_lin, for(age) ref(age 50) ci(hr_age_lin_lci hr_age_lin_uci) eform
Plots the results and interpret.

(e) Calculate a new set of Martingale-like residuals and plot vs age with a lowess smoother.
What does this tell you about the assumption of linearity in your model?
(f) Now use restricted cubic splines splines to model the effect of age. We will initially
use 4 df. We will store the locations of the knots and the projection matrix for the
orthogonalization for future use.
. rcsgen age, gen(rcsage) df(4) orthog
. matrix Rage = r(R)
. global knotsage r(knots)
. stpm2 rcsage1-rcsage4, scale(hazard) df(5) bhazard(rate)
Calculate a third set of Martingale-like residuals and plot vs age. Have the splines
captured the non-linearity in the data?
(g) The parameters in the model are difficult to interpret individually, but we can still get
useful predictions. Use the following code to obtain the predicted excess mortality
rates and predicted relative survival functions for subjects aged 40, 60 and 80 at
diagnosis.
. foreach age in 40 60 80 {
rcsgen , scalar(age) rmatrix(Rage) gen(c) knots(\$knotsage)
. predict hage, hazard ///
at(rcsage1 =c1 rcsage2 =c2 rcsage3 =c3 rcsage4 =c4) ///
timevar(temptime) per(1000)
. predict sage, survival ///
at(rcsage1 =c1 rcsage2 =c2 rcsage3 =c3 rcsage4 =c4) ///
timevar(temptime)
}
82 EXERCISES
Note the use of the rcsgen command. We need to know the value of the spline
variables for the ages of interest so we make sure we use the same knot locations and
projection matrix.
Plot the predicted excess hazard and relative survival functions. Which age group is
the most different?
(h) Obtain predicted 1-year relative survival with 95% confidence intervals as a function
of age using the following code.
. gen t1 = 1
. predict s1, survival timevar(t1) ci
Plot this function (vs age) and interpret.

(i) Repeat the above, but now obtain predicted 5-year relative survival as a function of
age.
(j) Calculate the 5-year relative survival conditional on survival to 1 year. (HINT Recall
that survival to time t conditional on survival to time t0 is S(t)/S(t0 )).
Plot the conditional relative survival versus age. Explain the similarity between 2-
year, 5-year and the conditional relative survival.
Note that if you want to obtain confidence intervals for conditional relative survival
you can use Statas predictnl command that uses the delta-method with numerical
derivatives.
. predictnl condsurv2 = predict(survival timevar(t5)) / ///
predict(survival timevar(t1)) ///
, ci(condsurv2_lci condsurv2_uci)
(k) Obtain the excess mortality rate ratio as a function of age at diagnosis using the
following code.
. rcsgen , scalar(50) rmatrix(Rage) gen(c) knots(\$knotsage)
. partpred hr_age_rcs, for(rcsage1-rcsage4) ///
ref(rcsage1 =c1 rcsage2 =c2 rcsage3 =c3 rcsage4 =c4) ///
eform ci(hr_age_rcs_lci hr_age_rcs_uci)
Plot this function (vs age) and compare to the graph you produced in part (d).
Interpret the graph and explain why it was important to model the effect of age at
diagnosis as a non-linear function.
(l) Optional Extra Question. Investigate the effect of using a different number of knots
to model the effect of age at diagnosis using restricted cubic splines. E.g. Try 3, 4
and 5 df using the rcsgen command. You can plot the different functions and also
compare the fit of the models using AIC or BIC.
83
232. Modelling non-linear effects in relative survival II Time-dependent effects
(a) Load the colon cancer data and merge in the background mortality rates as in ques-
tion 230. Remember to drop those aged over 90 years.
(b) Fit a proportional excess hazards flexible parametric relative survival model using
splines for the effect of age (see part (f) of question 231).
. stpm2 rcsage1-rcsage4, scale(hazard) df(5) bhazard(rate)
Use estimates store peh to save this model so we can later perform a likelihood
ratio test.
(c) Now fit a model with time-dependent effects for the (non-linear) effect of age. Use
2 df for the time-dependent effects. Perform a likelihood ratio test comparing this
model with proportional excess hazards model.
. stpm2 rcsage1-rcsage4, scale(hazard) df(5) bhazard(rate) ///
tvc(rcsage1-rcsage4) dftvc(2)
. estimates store timedep
. lrtest peh timedep
Is there evidence to reject the proportional excess hazards assumption?
(d) Predict the excess mortality rates and predicted relative survival functions for subjects
aged 40, 60 and 80 at diagnosis. Note that you can use exactly the same code
as part (g) of question 231 since the predict command of stpm2 will recognize
which covariates have been specified as having time-dependent effects. Compare
these graphs with those obtained in question 231. Are the functions different?
(e) Obtain the 1 year relative survival as a function of age and plot. You can use ex-
actly the same code as part (h) of question 231. Compare the estimates from the
proportional excess hazards model and the model allowing time-dependent effects.
(f) Obtain the 5 year relative survival as a function of age and plot. You can use ex-
actly the same code as part (i) of question 231. Compare the estimates from the
proportional excess hazards model and the model allowing time-dependent effects.
(g) Obtain the 5 year relative survival conditional on survival to 1 year as a function of age
and plot. You can use exactly the same code as part (j) of question 231. Compare
the estimates from the proportional excess hazards model and the model allowing
time-dependent effects. Why is the shape different to that obtained in question 231?
(h) We can still summarize our data as excess mortality rate ratios, but now not only is
the effect of age non-linear, the excess mortality ratios vary as a function of follow-up
time. Using age 50 as the reference age calculate the time-dependent excess mortality
rate ratios for subjects aged 40, 60, 70 and 80. The following code will help.
. rcsgen , scalar(50) rmatrix(Rage) gen(ref) knots(\$knotsage)
. foreach age in 40 60 70 80 {
rcsgen , scalar(age) rmatrix(Rage) gen(cage_) knots(\$knotsage)
predict hrage, ///
hrnum(rcsage1 =cage_1 rcsage2 =cage_2 ///
rcsage3 =cage_3 rcsage4 =cage_4) ///
hrdenom(rcsage1 =ref1 rcsage2 =ref2 ///
rcsage3 =ref3 rcsage4 =ref4) ///
timevar(temptime) ci
}
Interpret the excess mortality rate ratios.
84 EXERCISES
(i) By adapting the code in part (h) obtain differences the excess mortality rate for the
same ages with age 50 as the reference age. Plot the results and think about how the
relative effects seen in part(h) relate the absolute effects seen here.
Adapt the code once more to obtain differences in relative survival as a function of
follow-up time for the same ages with age 50 as the reference age.
(j) Perform a sensitivity analysis for the number of degrees of freedom you use to model
the time-dependence, i.e. using the dftvc option. Try between 1 and 3 df and predict
the time-dependent excess mortality rate ratio comparing a subject aged 70 with a
subject aged 50. Is the time-dependent effects sensitive to the number of df? Unless
you are an experienced Stata programmer you may want to look at the solution Do
file to help you.
85
240. Localised melanoma: age-standardised estimates of relative survival (for a sin-

gle cohort using an internal standard)
(a) By calculating a single life table with 1 year intervals, estimate the 10-year relative
survival ratio (Ederer II method) for all patients diagnosed with localised melanoma
1975-1994.
(b) Now calculate an age-standardised estimate of the 10-year RSR (Ederer II) using
traditional direct standardisation using an internal standard without using the stan-
dardisation options in strs. That is, calculate a life table for each age group to obtain
age-specific estimates and make a weighted average of these estimates (use the pro-
portion in each age group as weights) to obtain the age-standardised estimate. You
might wish to use MS Excel to help you perform the calculations where the worksheet
might contain the following cells.
agegrp cri wi cri wi
044
4559
6074
75+
Sum
The age standardised estimate is then calculated as i (cri wi )/ i wi .

P P
Does the age-standardised estimate differ from the crude estimate? Did you expect
it to differ?
(c) Now recalculate the age-standardised 10-year RSR (traditional direct standardisa-
tion) using the standardisation options in strs. Verify that you get the same answer.
HINT: Specifying the standstrata() option results in strs first producing stratified
life tables for each level of the variables specified in standstrata() and then pro-
ducing standardised estimates using the weights contained in the variable specified
in the iweights() option. We therefore need to first generate a variable containing
the weights. The weights must be specified as proportions so we will use the number
first at risk in each age stratum divided by the total number first at risk as weights.
. local totalobs = _N
. bysort agegrp: gen standwei = _N/totalobs
. strs using popmort [iw=standwei], br(0(1)15)
mergeby(_year sex _age) standstrata(agegrp)
(d) Now calculate the age-standardised 10-year RSR (Ederer II) using the Brenner al-
ternative method [17] using the standardisation options in strs. The approach is
identical to the previous part except we specify the brenner option.
Does the age-standardised estimate (alternative method) differ from the crude esti-
mate? Did you expect it to differ?
(e) In question 201g we estimated and compared cumulative relative survival using three
alternative methods for estimating expected survival (Hakulinen, Ederer I, and Ed-
erer II). Repeat this exercise but standardise by age (using traditional direct stan-
dardisation with an internal standard).
How does standardisation affect the magnitude of the differences between the 3 meth-
ods? Is this what you expected?
86 EXERCISES
(f) Now calculate the Pohar Perme estimate of relative survival. As this estimate does
not need to be estimated separately in each age group, do not include the by(agegrp)
option. You should make sure to split the time-scale into monthly intervals.
Compare the Pohar Perme estimates with the other estimates.
Plot the Pohar Perme estimates together with the Ederer II estimates.
87
241. Localised melanoma: age-standardised comparisons of relative survival be-

tween two periods of diagnosis
The aim of this exercise is to compare 10-year relative survival (Ederer II) between the
two calendar periods of diagnosis (19751984 & 19851994) whilst standardising for age.
We will use both the traditional and the alternative methods for age standardisation. The
age distribution for the first calendar period will be used as the standard.
(a) Crude estimates. Use strs to estimate the 10-year relative survival ratio (Ederer II
method, 1-year intervals) for each of the two calendar periods of diagnosis (19751984
& 19851994).
(b) Traditional method by hand. Without using the standardisation options of strs,
estimate age-standardised 10-year relative survival for patients diagnosed 19751984
(i.e., the early period) using the internal age distribution as the standard. That is,
repeat exercise 240b with the only difference being that we are restricting to the first
period of diagnosis.
Now estimate the age-standardised 10-year relative survival for patients diagnosed
19851994 (i.e., the latter period) using the age distribution of the early period as
the standard.
i. Did age-standardisation have a large impact (compared to the crude analysis)?
ii. Based on the stratum-specific survival estimates and stratum-specific weights
would you expect age-standardisation to have a large effect on the comparisons?
iii. Under what conditions would you expect the comparison of age-standardised
estimates to differ from the comparison of crude estimates?
(c) Traditional method using iweights option in strs. Repeat the previous part
using the iweights option in strs and confirm you get the same results.
(d) Alternative method using iweights and brenner options in strs. Now re-
peat the exercise using the alternative method of standardisation rather than the
traditional method. How does the choice of method for standardisation affect the
inference.
(e) Pohar Perme estimate. The Pohar Perme estimate of relative survival is an inter-
nally standardized estimate. Explain the dangers of comparing the crude estimates
between the two calendar periods.
Obtain the age standardized estimates of the Pohar Perme estimate of relative survival
in the two calendar periods using 19751984 as the reference age.
How do these compare to the Ederer II estimates?
88 EXERCISES
242. Age standardization using flexible parametric models
This question uses stpm2 to obtain model based age standardized estimates of relative sur-
vival. The results should be compared to the life table based estimates from questions 240
and 241.
This question assumes familiarity with using flexible parametric models to model excess
mortality and does not provide complete details of the required Stata code; we therefore
suggest you look at question 230 before attempting this question if you are unfamiliar with
flexible parametric models.
(a) Load the melanoma data, keep those with localized melanoma, merge in the expected
mortality at death/censoring and fit a flexible parametric relative survival model with
no covariates using 5 df for the baseline.
Create a temporary time variable.
Predict the all age combined relative survival curve (use the option timevar(temptime)).
What is the estimated relative survival at 10 years post diagnosis? How does this
compare to question 240(a).
(b) Fit a proportional excess hazards model for age group. Use the predict command to
estimate a relative survival curve for each age group. Plot these on a graph together
with the all age estimate to compare them.
(c) Use tab agegrp to calculate the the proportion of subjects in each age group. Calcu-
late a weighted average of the four age specific relative survival curves to obtain the
age standard relative survival curve. Note this is traditional direct standardization
using an internal standard. However, now the relative survival curves are model-based
rather than calculated separately in life tables.
Add the age standardized curve to the figure produced in part (b).
(d) Compare the 10 year age standardized estimate of relative survival to that obtained
in question 240.
(e) Calculate age-standardized relative survival using the meansurv option.
. predict rs_stand2, meansurv timevar(temptime)
Check that the estimate is the same as the calculated in part (c)
(f) You can calculate confidence intervals for the standardised estimate using the ci
option. This uses the delta method (with numerical derivatives) and will be slow if
you do not use the timevar option.
. predict rs_stand3, meansurv timevar(temptime) ci
Compare the confidence interval of the age standardized relative survival to that ob-
tained in question 240.
We now compare two periods of diagnosis in a similar way to question 241. If the age
distribution has changed then we need to standardize so that both time periods are
forced to have the same age distribution. As for question 241, the age distribution
for the first calendar period will be used as the standard
(g) Fit a proportional excess hazards model for age group and calendar period. Predict
and compare the 10 year relative survival in the eight groups.
89
(h) Is there evidence that the age distribution has changed between the two calendar
periods? If so, how has it changed?
(i) Use the meansurv option to obtain the age standardized estimate of relative survival
for each calendar period with the first calendar period as the reference.
. predict rs_7584 if year8594 == 0, meansurv at(year8594 0) timevar(temptime)
. predict rs_8594 if year8594 == 0, meansurv at(year8594 1) timevar(temptime)
Plot these on a graph and compare the 10 year age standardized relative survival to
the estimates in question 241.
(j) Now obtain the age-standardized estimate for 19851994 with this same calendar
period as the reference. Add this to the graph in (i). Explain the differences between
the curves.
90 EXERCISES
243. Localised melanoma: age-standardised estimates of relative survival (for a sin-

gle cohort using an external standard)
This question is similar to question 240, but here we will standardise to an external popu-
lation. For external standardisation it is common practice to standardise to a international
standard population. There are different standard populations depending on cancer site
(see [18].) The International Cancer Survival Standard (ICSS) for wide age-groups are
given in the table below. Weights are also available for 5-year age groups.
Weights
Age Group ICSS 1 ICSS 2 ICSS 3
044 0.07 0.28 0.60
4554 0.12 0.17 0.10
5564 0.23 0.21 0.10
6574 0.29 0.20 0.10
75+ 0.29 0.14 0.10
(a) Calculate the age-standardised 5-year RSR (traditional direct standardisation - Ed-
erer II method) using the standardisation options in strs for all patients diagnosed
with localised melanoma 1975-1994. Use the age groups defined in the table above.
HINT: to create the age-groupings and the internal weights to internally standardise,
you can use the code below.
.recode age (min/44=1) (45/54=2) (55/64=3) (65/74=4) (75/max=5), gen(agegrpICSS)
. label define agegrpICSS 1 "0-44" 2 "45-54" 3 "55-64" 4 "65-74" 5 "75+"
. label values agegrpICSS agegrpICSS
. label variable agegrpICSS "Age groups for ICSS"
. local totalobs = _N
. bysort agegrpICSS: gen standwei = _N/totalobs
. label variable standwei "Internal age group weights"
(b) For melanoma, we use the International Cancer Survival Standard (ICSS) 2 weights.
Calculate the externally age-standardised 5-year RSR using the standardisation op-
tions in strs by using the ICSS 2 weights given in the table above.
(c) Compare the estimates using the two different weights. Are they similar? Did you
expect them to be?
HINT: look at the values for the two weights for the 5 specified age groups.
(d) Repeat part (b) using the ICSS 1 weights instead.
What do you expect to happen to the standardised estimate when standardising to
an older age distribution?
91
250. Probability of death in a competing risks framework (life table relative sur-
vival)
strs implements the approach proposed by Cronin and Feuer (2000) [19] for estimating
the crude probability of death based on life table estimates of relative survival. We explore
the life table approach in this question. Lambert et al. (2010) [20] subsequently showed
how the estimates can be obtained after fitting a relative survival model, namely a flexible
parametric models for relative survival, which use restricted cubic splines for the baseline
cumulative excess hazard and for any time-dependent effects. The approach using flexible
parametric models for relative survival is covered in question 251. Although the two
approaches estimate the same quantity, the life table approach provides estimates for
grouped data so we get an estimated probability for an age group rather than an estimate
for a specific age as can be obtained in the model-based approach.
(a) Load the Melanoma data, drop subjects diagnosed 19751984 and then and use strs
to obtain life-tables stratified by age group and sex. Use the cuminc option to obtain
the crude probabilities of death due to cancer and due to other causes.
(b) How is the probability of death due to all causes, F, calculated?
(c) Why is the crude probability of death due to cancer, ci dc similar to the all-cause
probability of death for subjects aged 0-44?
(d) For both males and females aged 60-74 what is the probability of death due to all-
causes at 5 years post diagnosis? What two variables can be added together to give
the probability of death due to all-causes?
(e) What proportion of the all-cause deaths at 5 years post diagnosis are due to cancer
and due to other causes for males? Compare these figures for the different age groups.
(f) The age groups are fairly wide, explain how you would expect the crude probability
of death due to cancer to differ between a 60 and 74 year old, even if the relative
survival was identical.
(g) Plot the net probability of death, the crude probability of death due to cancer and the
overall probability of death for males by age group. Try to understand the relationship
between these various measures.
92 EXERCISES
251. Probability of death in a competing risks framework (relative survival model)

In exercise 250 we explored how one could estimate crude probabilities of death based on life
table estimates of relative survival making use of the strs implementation of the approach
proposed by Cronin and Feuer (2000) [19]. Lambert et al. (2010) [20] subsequently showed
how the estimates can be obtained after fitting a relative survival model, namely a flexible
parametric models for relative survival, which use restricted cubic splines for the baseline
cumulative excess hazard and for any time-dependent effects. Although the two approaches
estimate the same quantity, the life table approach provides estimates for grouped data
so we get an estimated probability for an age group rather than an estimate for a specific
age as can be obtained in the model-based approach.
(a) Load the Melanoma data and merge in the background mortality rates as in ques-
tion 230. Fit a flexible parametric relative survival model including age group with
time-dependent effects.
. stpm2 agegrp2-agegrp4, scale(hazard) bhazard(rate) df(5) ///
tvc(agegrp2-agegrp4) dftvc(3)
Calculate the estimated net mortality (1 - relative survival) and plot the four curves
on a single graph. Interpret the plot.
(b) Use the stpm2cm command to estimate the crude probability of death. Note that
stpm2cm will predict for individual covariate patterns and for ages at diagnosis. Per-
form the predictions for males aged 40, 55, 70 and 80 diagnosed in 1985. The predic-
tion for a 40 year old (the first age group) can be obtained using,
. stpm2cm using popmort, at(agegrp2 0 agegrp3 0 agegrp4 0) ///
mergeby(_year sex _age) ///
diagage(40) diagyear(1985) ///
sex(1) stub(cm1) nobs(1000) ///
tgen(cm1_t)
Plot the estimated crude probability of death due cancer for each of the selected ages
on the same graph. Contrast these with the estimated net probability of death from
part (a).
(c) Generate a similar plot but for the crude probability of death due to other causes.
(d) A useful way of presenting crude probabilities is through stacked graphs. Generate
the stacked graphs for each of the selected ages. Use the solution Do file for help.
(e) Advanced: Now fit a model using splines for the effect age with the spline terms
allowed to be time-dependent. Calculate the crude probabilities of death and compare
these to the model where age is categorized.
93
260. Fitting cure models

Stata addon required! This exercise requires the Stata user-written command strsmix.
We will now apply cure fraction models [21, 22] to the colon cancer data. In this exercise
we fit mixture cure models and in exercise 261 we fit flexible parametric cure models. The
cure fraction models treat time as continuous and thus there is no need to split the time
scale. However, the expected hazard (mortality) rate at the time of death (or censoring)
is required. Use the following commands to merge in the expected mortality rate.
. use colon
. stset surv_mm, failure(status=1 2) scale(12) exit(time 120)
. merge m:1 _year sex _age using popmort, keep(match master)
The scale(12) option converts survival time to years. The exit(time 120.5) option
creates a maximum follow-up time of 10 years (120 months).
(a) Explain the purpose of the two gen statements in the above stata code.
(b) Fit a mixture cure fraction model to those diagnosed between 1975-1984 using the
following command.
. strsmix if year8594==0, dist(weibull) link(identity) bhazard(rate)
i. What is the estimate of the cure fraction?

Use the following commands to obtain prediction of the relative survival curve and
the survival distribution of the uncured and then plot these estimates against
time (_t)
. predict rs7584, survival
. predict rs7584u, survival uncured
ii. Does the relative survival curve appear to reach a plateau at the cure fraction?
Would you expect it to?
iii. Approximately what proportion of the uncured group have died after 2 years?
iv. Approximately what is the median survival time of the uncured ?
(c) Repeat the above for those diagnosed between 19851994. Contrast the estimates for
the two time periods.
(d) Now we will compare the two time periods more formally by including (year8594)
as a covariate. First just allow the cure fraction to vary between time periods.
. strsmix year8594, dist(weibull) link(identity) bhazard(rate)
i. What is the estimated difference in the cure fraction between the two time peri-
ods? Contrast this to the estimates obtained in b(i) and (c).
94 EXERCISES
ii. This model is making a fairly strong assumption regarding the survival distribu-
tion of the uncured for the two periods. What is this assumption?
Now allow the two Weibull parameters ( and ) to vary between the two time
periods.
. strsmix year8594, dist(weibull) link(identity) bhazard(rate)
k1(year8594) k2(year8594)
iii. What is the estimated difference in the cure fraction between the two time peri-
ods? Contrast this with d(i).
iv. Test the assumption that the survival distribution of the uncured is the same
for the two time periods.
(e) Now fit a model including age group and time period of diagnosis using a logit link
(use option link(logit)).
i. Interpret the parameter estimates (you may want to display the exponentiated
coefficents bys using strsmix, eform).
ii. Obtain predictions of the median survival of the uncured.
Hint, use predict med, centile to obtain predicted values of the median.
95
261. Fitting cure models using flexible parametric survival models

We will now apply flexible parametric cure models to the same data as in exercise 260,
where we fitted mixture cure models. Read in the data, stset and merge on expected
mortality rates in the same way as in exercise 260.
(a) Compare the cure proportion in the two time periods by including the variable
year8594 as a covariate in the stpm2 command. Assume proportional hazards.
. stpm2 year8594, df(6) bhazard(rate) scale(hazard) cure
i. How do you interpret the coefficient for the effect of the time period?
ii. Use the coefficients in the output to calculate the estimated cure proportions for
iii. Predict the cure proportions using the predict command to check your calcula-
tions.
. predict cure1, cure
. list cure1 if year8594==0, constant
iv. What is the estimated difference in the cure proportion between the two time
periods? Compare this to the estimates obtained in exercise 260. Are the results
similar? Would you expect them to be similar?
v. Predict the median survival time of uncured. Is the median survival time the
same in the two groups? Should it be?
. predict med1, centile(50) uncured
. list med1 if year8594==0, constant
(b) Now allow time-dependent effect.

. stpm2 year8594, df(6) tvc(year8594) dftvc(4) bhazard(rate) scale(hazard) cure
i. How do you interpret the coefficient for the effect of the time period?
ii. Use the coefficients in the output to calculate the estimated cure proportions for
iii. Predict the cure proportions using the predict command to check your calcula-
tions.
. predict cure2, cure
iv. Are the cure proportions similar to what was estimated in (a)?
96 EXERCISES
v. Predict the median survival time of uncured. Is the median survival time the
same in the two groups? Should it be? Is the difference between the periods
smaller or larger than in (a)? Why?
. predict med2, centile(50) uncured
(c) Plot the estimated overall relative survival and the relative survival among uncured
for the two periods. Do the survival curves reach a plateau? Should they?
97
280. Constructing a popmort file for Stata using data from the Human Mortality
Databases
The Human Mortality Database (HMD) was created to provide detailed mortality and
population data to researchers, students, journalists, policy analysts, and others interested
in the history of human longevity. The project began as an outgrowth of earlier projects
in the Department of Demography at the University of California, Berkeley, USA, and
at the Max Planck Institute for Demographic Research in Rostock, Germany. The HMD
currently maintains data from 37 different countries. In this exercise we will provide a
step-by-step introduction to how country-specific popmort files can be downloaded from
the HMD web site and converted into Stata format.
The data in the HMD project are provided free of charge to all individuals who request
access to the database. However, before gaining full access to the database, you must
become a registered user, which requires accepting our user agreement and answering just
a few questions.
(a) Go to the HMD web-page http://www.mortality.org and complete the registration

form for new users. After having completed the registration you will receive an e-mail
with your user name and password.
(b) In this exercise we will create a popmort file for Sweden. In order to be directed to
the Swedish raw data, go to the main home page and click on Sweden.
(c) Several types of data sets, with different levels of resolution, are generally available for
each country. The country-specific data sets are documented by HMD and the meta
data is available from the web-page. To create a popmort file of the same structure as
those used throughout the course click on the cell that contains 1 x 1 period estimates
of the death rates. You will be asked to enter your username and password to open
the text file that contains these rates.
98 EXERCISES
(d) Look through the data file and make sure that you understand the structure of this
file. The first few lines are reproduced below. Download the text file to your computer
and save it at some convenient location. Note that you will, in a later step, need to
type the file reference so saving it on the desktop may not be the best option.
Sweden, Death rates (period 1x1) Last modified: 23-Aug-2012,(May07)
Year Age Female Male Total

1751 0 0.212235 0.241105 0.226774
1751 1 0.049412 0.052949 0.051169
1751 2 0.032247 0.034587 0.033409
1751 3 0.026006 0.027883 0.026936
1751 4 0.023696 0.025692 0.024681
1751 5 0.018761 0.020801 0.019766
1751 6 0.012956 0.014635 0.013788
(e) Before the data can be read into Stata you need to make sure that Stata will be able
to recognize all numeric variables. In the Swedish data the death rates for people
older than 109 years are denoted as 110+. Use the find and replace function (of a text
editor or the Stata do file editor) to replace 110+ with 110. Also, remove the first
line of text and the second blank line in the text file (but keep the variable names).
(f) Read the data into Stata using the infile command (see below). Note that we
restrict the file to years 1950 onwards and ages below 100.
. infile _year _age female male total using "Name of file.txt " ///
if (_year > 1949 & _age < 100), clear
(g) The rates for males and females are stored in separate columns. In the final popmort
file, we would like to have the rates for males and females stacked on top of each other
and a variable called sex that enables us to separate the rates. This can be achieved
by running the following code:
. drop total
. rename male rate1
. rename female rate2
. reshape long rate, i(_year _age)
. rename _j sex
(h) Next, we will use the mathematical relationship between mortality rates and survival
probabilities to create a new variable, prob, that contains the survival probabilities
for each covariate pattern (in addition to the mortality rates that were available in
the raw data file).
. gen prob=exp(-rate)
(i) The popmort file is now ready to be used in survival analysis. The following command
can be used to add some final touches to the file. Do not forget to save the cleaned
popmort file.
. label data "Swedish death rates from http://www.mortality.org/"
. label variable rate "Death rate"
. label variable prob "Survival probability"
. label variable _year "Year of death"
. label variable _age "Age"
. label variable sex "Sex"
. save popmort2011, replace
Voila!
99
281. Constructing a popmort file by modelling cohort data
This code illustrates how one can create a popmort file using data from a cohort. If
you are interested in a standard popmort file then the first place to look is the Human
Mortality Database or your national statistics office. The approach described here is for
when those approaches dont suffice. For example, if you wish to create a popmort file
stratified by region or social class and you have access to a cohort of individuals followed
up for death.
In this illustration we use data on cancer patients (because that is what we have) and
use non-cancer mortality as the outcome. In a real life appplication, however, you would
ideally have data on individuals from the general population and use all-cause mortality
as the outcome.
. use colon, clear
/* We need a variable for date of birth */

. gen dob=dx-age*365.25
. format dob %d
/* stset using attained age as the timescale */

. stset exit, fail(status==2) enter(dx) origin(dob) scale(365.25) id(idnr)
/* graph the age-specific hazards for each sex */

. sts graph if _t < 100, haz by(sex) name(observed, replace) kernel(epan2) ///
xscale(range(40 100)) xlabel(40(5)100) ///
yscale(range(0 1.5)) ylabel(0(0.5)1.5)
/* Tabulate age-specific mortality rates */

. preserve
. stsplit attage, at(0(1)110)
. strate sex attage
. restore
Model the age-specific hazards for each sex. First using a proportional hazards model.
. stpm2 sex, df(5) scale(hazard)
/* Predict the hazards and plot by sex*/

. predict h, haz
. twoway (line h _t if sex==1 & _t>40,sort) ///

(line h _t if sex==2 & _t>40, sort), ///
name(fitted, replace) title("Fitted values from fpm (prop hazards)") ///
yscale(range(0 1.5)) ylabel(0(0.5)1.5)
/* Plot the empirical and fitted hazards (FPM PH model) together */

. sts graph if _t < 100, haz by(sex) kernel(epan2) name(overlay1, replace) ///
title("Empirical and fitted hazards (fpm PH)") ///
addplot(line h _t if sex==1 & _t>40,sort || ///
line h _t if sex==2 & _t>40, sort)
100 EXERCISES
/* Same plot but on the log scale */

. sts graph if _t < 100, haz by(sex) kernel(epan2) name(overlay2, replace) ///
title("Empirical and fitted hazards (fpm PH)") ///
yscale(log) ylabel(0 0.01 0.1 0.5 1.5) ///
addplot(line h _t if sex==1 & _t>40,sort || ///
line h _t if sex==2 & _t>40, sort)
Now allow non-proportional hazards.
. stpm2 sex, df(5) scale(hazard) tvc(sex) dftvc(3)

. predict h2, haz
. sts graph if _t < 100, haz by(sex) kernel(epan2) name(overlay1_tvc, replace) ///
title("Empirical and fitted hazards (fpm tvc)") ///
addplot(line h2 _t if sex==1 & _t>40,sort || ///
line h2 _t if sex==2 & _t>40, sort)
. sts graph if _t < 100, haz by(sex) kernel(epan2) name(overlay2_tvc, replace) ///
title("Empirical and fitted hazards (fpm tvc)") ///
yscale(log) ylabel(0 0.01 0.1 0.5 1.5) ///
addplot(line h2 _t if sex==1 & _t>40,sort || ///
line h2 _t if sex==2 & _t>40, sort)
Create a table of predicted rates and probabilities. These figures could be used as the
popmort file for relative survival analysis. For a real-life example we would also stratify
by calendar year and race (if applicable). Could also model rates as a function of social
class or ethnicity. To estimate rates for a small population one could include data for a
larger population to get the shape correct.
. range attage 40 100 61

. drop if attage==.
/* Use non PH model to predict hazards for these ages (both sexes)*/
.predict rate_m, hazard at(sex 1) timevar(attage)
. predict rate_f, hazard at(sex 2) timevar(attage)
/* Generate survival probabilities*/

. gen prob_m=exp(-rate_m)
. gen prob_f=exp(-rate_f)
. format rate_m rate_f prob_m prob_f %8.5f

. keep attage prob_m prob_f rate_m rate_f
. list
101
282. Calculating excess and avoidable deaths from life tables
(a) Load the Melanoma data, drop subjects diagnosed 1975-1984 and then and use strs
to obtain life-tables stratified by age group and sex. Load the grouped data and keep
the following variables.
. keep start end n cp cp_e2 cr_e2 sex agegrp
(b) What is the difference in five-year relative survival between males and females in each
age group?
(c) We will now investigate excess deaths and avoidable deaths. The question of interest
is how many fewer deaths we would expect to see if males could achieve the same
relative survival as females. To do this we will reshape the data from long form to
wide form to make calculations easier.
. bysort sex (agegrp start): gen j = _n
. gen sexlab =cond(sex==1,"_m","_f")
. drop sex
. reshape wide start end n cp cp_e2 cr_e2 agegrp, i(j) j(sexlab) string
. rename agegrp_m agegrp
. rename start_m start
. rename end_m end
. drop agegrp_f start_f end_f
Look at the data in the data browser to make sure you understand what the reshape
command has done.
(d) In order to calculate the predicted number of deaths we need to define how many
subjects were at risk at the the start of follow-up. For simplicity, we will use the
average number of cases per year over the 10 year diagnosis period. This can be
calculated as follows.
. bys agegrp: gen Nrisk_m = n_m[1]/10
Calculate the overall (all-cause) probability of death, 1 S (t)R(t), for males.
. gen p_dead_m = 1 - cp_e2_m * cr_e2_m
For males, calculate the expected number of all-cause deaths, Nd m, the expected
number of deaths if the study population were free of cancer, NExp d m and the excess
deaths associated with a diagnosis of cancer, ED m.
. gen Nd_m = Nrisk_m*p_dead_m
. gen NExp_d_m = Nrisk_m*(1-cp_e2_m)
. gen ED_m = Nd_m - NExp_d_m
i. How many all cause deaths would we expect to see in each age group at 5 years
post diagnosis?
ii. How many more deaths are there than would be expected in a similar cancer free
group in the population?
iii. How many excess deaths by 5 years are associated with a diagnosis of melanoma
over all age groups?
(e) Repeat the above calculations for females. How do the excess deaths for females
compare to the males?
102 EXERCISES
(f) We will now apply the relative survival estimates for females to the males expected
survival in order to calculate the avoidable deaths.
. gen Nd_m_f = Nrisk_m*(1 - cp_e2_m * cr_e2_f)
. gen AD_m = Nd_m - Nd_m_f
How many deaths would be avoided if males could achieve the same relative survival
as females for Melanoma?.
(g) List the avoidable deaths for the oldest age group over all follow-up times. Why are
the number of avoidable deaths decreasing as follow-up time increases?
103
283. Simulating relative survival
In this exercise, the aim is to learn how to simulate relative survival data. We will go
through this step-by-step and show the process in a fairly simple setting. To carry out
the exercises you will need to install some user-written commands from within Stata. The
survsim command is used to simulate survival data from a range of distributions in Stata.
. ssc install survsim
The process of generating relative survival data through simulation involves simulating
two times of death from differing mortality rates. Firstly, we generate the cause-specific
time of death; when patients would die in the hypothetical scenario where they can only
die of the disease of interest. Secondly, we generate the time of death that patients would
experience if they were only at risk of other causes of death (based on general population
mortality rates). We then take the minimum of these two times to be our all-cause time
of death.
(a) Firstly, use the code below to set the observations to 1000 and create an age and year
of diagnosis variable. Here, we use a normal distribution for age with a mean of 60
and a standard deviation of 13. This means that age-at-diagnosis will be a random
draw from a normal distribution for these 1000 patients. We set year of diagnosis
to 1990 for all patients and make them all males for simplicity. We also create an
age-group variable using the standard 5 age-groups.
. clear
. set obs 1000
. gen agediag=floor(rnormal(60,13))
. gen yydx=1990
. gen sex=1
. egen agegrp=cut(agediag), at(0 45 55 65 75 110) icodes
Plot a histogram for age to see the distribution for your sample.
(b) We now need to use survsim to generate cause-specific survival times for our patients.
These times will relate to times of death due to the disease of interest only meaning
that we will calculate the time that the patient would have died of their disease for
all patients. We will assume that the survival times follow a Weibull distribution
with = 0.2 and = 0.5. Use the code below to generate the survival times. We
will also assume a proportional effect for categorised age. The hazard ratios for the
5 groups will be 0.8, 0.9, 1, 1.2 and 1.4 respectively. Remember that the baseline will
be defined by the and above.
. survsim timerel, lambdas(0.2) gammas(0.5) ///
covariates(agegrp1 =ln(0.8) agegrp2 =ln(0.9) agegrp3 =ln(1) ///
agegrp4 =ln(1.2) agegrp5 =ln(1.4))
Use twoway function to plot the survival function for the 5 age-groups.
HINT: twoway function y=exp(-0.2*x^0.5), range(0 5) will plot the Weibull
survival curve for the baseline age-group.
(c) Now we have the cause-specific times of death, we also need the time of death due
to other causes for each patient. Using both these times, we can create an all-cause
death time variable for our stset. We will use the population mortality file to draw
survival times for our patients (assuming an exponential distribution); we will need to
do this a number of times because our patients will age over follow-up and therefore
their mortality rates will change as they age. Therefore, for each year of follow-up;
104 EXERCISES
we will calculate an expected survival time and evaluate if the patient survives the
year. We will let them age by a year and recalculate the times based on the new
rates. Here, we do this 5 times so that we have a minimum of 5 years of follow-up.
. gen _age=.
. gen _year=.
. forvalues i=0/4 {
capture drop _merge prob
replace _age=agediag+i
replace _year=yydx+i
quietly merge m:1 _age _year sex using popmort, keep(matched master)
gen timebacki=(-log(runiform()))/(-ln(prob))
replace timebacki=1 if timebacki>=1
}
. gen timeback=.
. forvalues i=0/4 {
quietly replace timeback=timebacki+i if timebacki<1 & timeback==.
}
. drop timeback? _age _year _merge rate
timeback now contains the background survival time based on the population mortal-
ity rates. timeback is created from the five numbered timeback variables by adding
a year for each interval a patient survives, or stopping the sum once the patient dies
in any given interval.
(d) We now take the minimum of the cause-specific survival times and the background
survival times to obtain our all-cause survival time estimates. We censor patients at
5 years should they survive for 5 years or more for both time variables; creating a
death indicator in the process.
. gen time=min(timeback,timerel)
. generate died = time <= 5

. replace time = 5 if died == 0
. stset time, failure(died = 1)
We can now stset our data as though we have the all-cause data that we would have
in real-life. However, we now know the true values for the age-group specific survival
curves from our Weibull parameters. We also know the true distribution of age is
normal and so can calculate the proportion in each age-group using z-values. We can
also create an indicator variable indicating whether each patient died of cancer or
other causes; we have perfect cause of death information because we generated the
data. We could have just fitted a cause-specific analysis to the initial survival times
from survsim. However, it is also useful to be able to simulate data suitable for a
relative survival analysis; using stset on the all-cause data, so that we can evaluate
the properties of our relative survival estimation methods.
105
(e) Now that we have generated the survival data, we can treat this as any normal
dataset and perform the analysis that we wish to evaluate through simulation. As
usual, we now must merge in the background mortality data before performing an
excess mortality model for age using stpm2.
. gen _age = min(int(agediag + _t),99)
. quietly merge m:1 _age _year sex using popmort, keep(matched master)
. stpm2 ib2.agegrp, bhaz(rate) df(5) scale(hazard) ///

eform
Compare the estimated hazard ratios to those you used in the generation of the
survival data using survsim. They may well be quite different - this is a single run
of the simulation with quite a small sample size.
(f) We now have all of the components to create a Stata program to perform a simulation.
We can run the simulation 1000 times, for instance, to evaluate the unbiasedness of
our parameter estimates for the flexible parametric excess mortality model. See the
end of the solution do file for how this can be done.
. simulate agegrp0=r(agegrp0) agegrp1=r(agegrp1) ///
agegrp2=r(agegrp2) agegrp3=r(agegrp3) ///
agegrp4=r(agegrp4), reps(1000): relsurvsim, hazratio(0.8 0.9 1 1.2 1.4)
106 EXERCISES
284. Estimating loss in expectation of life
In this exercise the aim is to estimate the loss in expectation of life for the melanoma
cohort as a function of age, year and sex. This can be used to estimate the total number
of life years lost for a given cohort of cancer patients. We will also use loss in expectation
of life as a way of quantifying the sex difference in melanoma survival, as an alternative
to using avoidable deaths (exercise 282).
Loss in expectation of life, together with life expectancy in absence of cancer and life
expectancy in presence of cancer can be estimated after fitting a flexible parametric model
by using the lifelost option of the predict postestimation command after using stpm2
to fit a model. All options used together with lifelost are described below:
mergeby(string) specifies the variables by which the file of general population

survival probabilities is sorted.
diagage(name) specifies the variable containing age at diagnosis.
Default is diagage.
diagyear(name) specifies the variable containing calendar year of
diagnosis. Default is diagyear.
maxage(int 99) specifies the maximum age for which general population survival
probabilities are provided in the using file. Probabilities
for individuals older than this value are assumed to be the same as
for the maximum age. Default is 99.
attage(name) specifies the variable containing attained age in the popmort file.
This variable cannot exist in the patient data file. Default is age.
attyear(name) specifies the variable containing attained calendar year in the popmort file.
This variable cannot exist in the patient data file.
Default is year.
survprob(name) specifies the variable containing survival probabilities in the popmort file.
This variable cannot exist in the patient data file. Default is prob.
using(string) specifies the popmort file to be used for expected survival probabilities.
by(string) specifies stratification variables. Survival probabilities are averaged for each
combination of these variables and assumed the same within each
combination. Can only be used together with the grpd option.
maxyear(int 2050) specifies the maximum age for which general population survival
probabilities are provided in the using file. Probabilities for years beyond
this value are assumed to be the same as for the maximum year.
Default is 2050.
nodes(int 50) specifies the number of nodes to be used for the numerical integration.
Default is 50.
tinf(int 50) specifies the end year used for the numerical integration. Both observed
and expected survival is assumed to be 0 after this point.
Default is 50.
tcond(real 0) specifies the starting year used for the numerical integration.
This is used to retrieve conditional estimates. Default is 0.
grpd specifies that average survival probabilities should be used, as opposed to
individual probabilities. If this is used together with the by option, the
average is calculated within each combination of the specified by variables.
stub(string) stubname for estimated life expectency in absence and presence of cancer.
107
(a) Load the melanoma data and stset the data for relative survival.
. gen patid = _n
. stset surv_mm, failure(status=1 2) scale(12) exit(time 120.5) id(patid)
(b) Fit a flexible parametric model including year, age and sex. Include age and year
as continuous variables using splines. Allow all covariates to have a time-dependent
effect. Remember to merge on the expected mortality at the exit times.
. rcsgen age, df(4) gen(sag) orthog
. rcsgen yydx, df(4) gen(syr) orthog
. gen fem= sex==2

. merge m:1 _year sex _age using popmort, keep(match master) keepusing(rate)
. drop _age _year _merge
. stpm2 sag1-sag4 syr1-syr4 fem, scale(hazard) df(5) ///

bhazard(rate) tvc(sag1-sag4 syr1-syr4 fem) dftvc(3)
(c) We will now estimate the loss in expectation of life. To save time we dont estimate
confidence intervals, although they can be obtained by removing the comments around
the ci option. (NOTE! Dont attempt to run this with the ci option during the lab
session. This would take more than an hour, and the only way to stop Stata is to
force the program to shut down completely.)
. predict ll, lifelost mergeby(_year sex _age) diagage(age) ///
diagyear(yydx) nodes(40) tinf(80) using(popmort) ///
stub(surv) maxyear(2000) /*ci*/
(d) Create a graph that shows how the loss in expectation of life varies over age, for males
diagnosed in 1994.
. twoway (line ll age if sex==1 & yydx==1994, sort) , legend(off) ///
scheme(sj) name(q41_d, replace) ytitle("Years", size(*0.8)) ///
xtitle("Age at diagnosis", size(*0.8)) xlabel(, labsize(*0.7)) ///
ylabel(0 5 10 15 20 25 30 35 40 45, labsize(*0.7) angle(0)) ///
yscale(range(0 45))
(e) List the life expectancy and the loss in expectaion of life for someone aged 50, 60, 70
and 80 at diagnosis, both males and females. Also calculate the total number of life
years lost among patients diagnosed in 1994.
. foreach age in 50 60 70 80
foreach sex in 1 2
list age sex yydx survexp survobs ll if age==age & ///
sex==sex & yydx==1994, constant
. qui summ ll if yydx==1994

. display r(sum)
108 EXERCISES
(f) Now estimate the loss in expectation of life if male patients had the same mortality
due to melanoma as female patients, but the expected survival of males.
. replace fem=1
. predict ll_alt, lifelost mergeby(_year sex _age) diagage(age) ///

diagyear(yydx) nodes(40) tinf(80) using(popmort) ///
stub(surv_alt) maxyear(2000) /*ci*/
(g) How many life years could potentially be saved if males diagnosed in 1994 had the
same survival from melanoma as female patients diagnosed in 1994?
. gen lldiff= ll-ll_alt
. summ lldiff if yydx==1994
. display r(sum)
. foreach age in 50 60 70 80
list ll ll_alt lldiff age if sex==1 & age==age & yydx==1994, constant
109
285. Multiple imputation for missing covariate data
This exercise gives an introduction to approaches for dealing with missing covariate data
in population-based cancer studies. Falcaro et al. (2015) [23] published a nice overview of
methods for modelling net survival when covariate data are missing.
In this exercise we use the official Stata commands for missing data that were introduced
with Stata release 12 (July 2011). Users of earlier Stata versions can perform the same
analyses with the user-written commands ice and mim.
Mechanisms for missingness in survival analysis can be classified as follows:
Missing completely at random (MCAR)

Covariate-dependent missing at random (CD-MAR)
Missing at random (MAR)
Missing not at random (MNAR)
Much of the literature on missing data uses three classifications (MCAR, MAR, and
MNAR). Falcaro et al. (2015) [23] distinguish between MAR (where the probability of
missingness depends on known covariates, the event indicator, and the survival time) and
CD-MAR (where the probability of missingness depends on known covariates, but not the
the event indicator or survival time).
The analytic approach depends on the mechanism. Multiple imputation requires MAR; if
the data are MNAR then the problem is more difficult. We cannot use the observed data
to identify the mechanism, we need substantive scientific knowledge of the processes that
gave rise to the data. The validity of inference depends on the (untestable) assumptions
inherent in the mechanism.
(a) Load the colon data and merge in the background mortality information as in ques-
tion 230. Tabulate stage. What proportion of observations have stage missing (i.e.,
unknown stage)?
(b) Investigate the distribution of unknown stage across age group and gender. Are older
patients more likely to have an unknown recorded stage?
(c) Study the estimated relative survival as a function of age group and stage (including
unknown).
stpm2 ib1.stage##i.agegrp , df(5) bhaz(rate) scale(hazard) eform nolog
predict survival, surv
line survival _t if stage==0, sort lpattern(dash) || ///

line survival _t if stage==1, sort || ///
line survival _t if stage==2, sort || ///
line survival _t if stage==3, sort ///
legend(order(1 "Unknown" 2 "Localised" 3 "Regional" 4 "Distant")) ///
by(agegrp)
Do you see anything notable about the survival of the patients with unknown stage?
110 EXERCISES
(d) Which mechanism do you feel is applicable for stage in these data? The following
questions are relevant to arrive at an answer.
i. Will the probability of missingness depend on survival time?
ii. Upon what known factors (if any) do you think missingness might depend?
iii. Do you think the probability of missingness might depend on factors that are not
available in our data? Which factors?
Note that this question (and the next) cannot be answered based on the observed data.
It requires knowledge of the process by which stage is assessed and recorded. Answer
the question based on your knowledge of such processes in your own jurisdiction. Talk
to a classmate or teacher if you dont possess extensive knowledge in this area.
(e) Multiple imputation requires an assumption of MAR (i.e., it is not valid for MNAR).
Do you think a MAR assumption is appropriate for the available data?
Note that we will proceed with multiple imputation, an approach that requires an
assumption that stage is MAR. Stata will provide estimates irrespective of what we
think about the MAR assumption.
(f) Using the code below, fit a flexible parametric model for excess mortality with ex-
planatory variables stage (localised as the reference category) and age group. For
pedagogic simplicity, we will not include a stage by age interaction, despite the fact
that the previous graph suggests it might be needed. This approach to analysing
incomplete data is known as the missing indicator approach (since individuals with
missing stage are included in a separate category).
. stpm2 ib1.stage i.agegrp, df(5) bhazard(rate) scale(hazard) eform
(g) Now change the coding of the variable stage so that the unknown category is coded
as missing (replace stage=. if stage==0). Refit the model from the previous
part, using the exact same code, and compare the number of observations and the
estimated excess hazard ratios. This approach is known as the complete records
approach for analysing incomplete data.
Both the missing indicator approach and the complete records approach are known to
be biased unless the data are missing completely at random (a scenario which rarely
occurs in practice). We will now explore an approach, multiple imputation, that is
preferred over the nave approaches.
111
(h) Before asking Stata to impute missing values we need to set the data for multiple
imputation (mi set) and declare those variables that contain missing values (stage
in this example).
mi set mlong
mi register imputed stage
Before performing the multiple imputation, give your best guess of what you think
the distribution of stage should be for each of the following three observations.
+------------------------------------------------------------------------+
| id agegrp sex subsite stage _t _d |
|------------------------------------------------------------------------|
| 2287 45-59 Female Coecum and ascending . .04166667 1 |
| 3362 75+ Female Coecum and ascending . 6.2083333 1 |
| 3501 75+ Female Coecum and ascending . 10 0 |
+------------------------------------------------------------------------+
That is, if you think each of the three stages are equally likely then specify 33.33%
for each. If you think the only possibility for stage is distant then specify 100% for
distant and 0% for the other two alternatives.
id agegrp _t _d localised regional distant
2287 45-59 0.04 1
3362 75+ 6.21 1
3501 75+ 10.0 0
We now multiply impute missing values using chained equations. Theory dictates
that the imputation model should contain the outcome. Recent research suggests
this be specified using the event indicator and estimated cumulative hazard. We
therefore generate the Nelson-Aalen estimate of the cumulative hazard and store it
in the variable H. We set the seed so as to obtain reproducible results (i.e., so you
get the same answers as in the solutions) but that step is not necessary in practice.
Carpenter and Kenward [24, p. 55] suggest 30 imputations but we will use only 10.
sts gen H=na
set seed 29390
mi impute chained (mlogit) stage = i.subsite sex i.agegrp H _d, add(10)
What type of model has been fitted for the imputation model? What covariates are
included in the imputation model?
(i) Examine the imputed values for the three observations we previously tried to predict
the distribution of stage. The variable _mi_m is zero for the observations in the
original data and a sequential integer for the imputed observations.
. list id _mi_m agegrp sex stage _t _d if id==2287
How closely did your predictions match the distribution of imputed values?
112 EXERCISES
(j) We now refit the flexible parametric model, this time to the imputed data. The mi
estimate command effectively fits the model to each of the 10 imputed data sets and
then combines the resulting estimates. Since stpm2 is not an official Stata command
we need to specify the cmdok option to specify that the command is OK for use with
imputed data. We also save the resulting estimates in order to make predictions in a
later step.
mi estimate, dots cmdok saving(mi_stpm2,replace): ///
stpm2 ib1.stage i.agegrp, df(5) bhaz(rate) scale(hazard) nolog
Now predict the survival function based on the fitted model.
mi predictnl survimp2 = predict(survival at(agegrp 2)) using mi_stpm2
(k) Using the fact that the original data is still in the new dataset (with _mi_m==0), we
can refit the model using the complete records approach and obtain predictions of
survival. We will graphically compare the predicted relative survival curves from the
imputation model and the complete records model for the 60 74 age-group.
stpm2 ib1.stage i.agegrp if _mi_m==0, df(5) scale(h) bhaz(rate)
predict surv, survival at(agegrp 2)
line surv survimp2 _t if stage==1 & _mi_m==0, sort || ///

line surv survimp2 _t if stage==2 & _mi_m==0, sort || ///
line surv survimp2 _t if stage==3 & _mi_m==0, sort ///
title("Predicted survival for agegrp==2 (60-74)") ///
legend(order(1 "Localised (Complete)" 2 "Localised (Imputed)" ///
3 "Regional (Complete)" 4 "Regional (Imputed)" ///
5 "Distant (Complete)" 6 "Distant (Imputed)")) ///
name(imputed, replace)
(l) If you want to try these methods further with this dataset then you can artificially
create missing data. For example, to randomly make agegrp missing for 25% of the
study population use,
. replace agegrp = . if runiform()<0.25
REFERENCES 113
References
[1] Dickman PW, Coviello E. Estimating and modelling relative survival. The Stata Journal
2015;15:186215.
[2] Lambert PC, Royston P. Further development of flexible parametric models for survival
analysis. The Stata Journal 2009;9:265290.
[3] Andersson TML, Lambert PC. Fitting and modeling cure in population-based cancer stud-
ies within the framework of flexible parametric survival models. The Stata Journal 2012;
12:623628.
[4] Lambert PC. Modeling of the cure fraction in survival studies. The Stata Journal 2007;
7:351375.
[5] Hinchliffe SR, Lambert PC. Extending the flexible parametric survival model for competing
risks. The Stata Journal 2013;13:344355.
[6] Clayton D, Hills M. Statistical Models in Epidemiology. Oxford: Oxford University Press,
1993.
[7] Royston P, Parmar MKB. Flexible parametric proportional-hazards and proportional-odds

models for censored survival data, with application to prognostic modelling and estimation
of treatment effects. Statistics in Medicine 2002;21:21752197.
[8] Volinsky CT, Raftery AE. Bayesian information criterion for censored survival models.
Biometrics 2000;56:256262.
[9] Sauerbrei W, Royston P, Bojar H, Schmoor C, Schumacher M. Modelling the effects of

standard prognostic factors in node-positive breast cancer. german breast cancer study
group (gbsg). British Journal of Cancer 1999;79:17521760.
[10] Ulm K. A simple method to calculate the confidence interval of a standardized mortality
ratio (SMR). American Journal of Epidemiology 1990;131:3735.
[11] Pohar Perme M, Stare J, Estève J. On estimation in relative survival. Biometrics 2012;
68:113120.
[12] Dickman PW, Sloggett A, Hills M, Hakulinen T. Regression models for relative survival.
Stat Med 2004;23:5164.
[13] Estève J, Benhamou E, Croasdale M, Raymond L. Relative survival and the estimation of
net survival: elements for further discussion. Statistics in Medicine 1990;9:529538.
[14] Hakulinen T, Tenkanen L, Abeywickrama K, Paivarinta L. Testing equality of relative

survival patterns based on aggregated data. Biometrics 1987;43:313325.
[15] Lambert PC, Smith LK, Jones DR, Botha JL. Additive and multiplicative covariate regres-
sion models for relative survival incorporating fractional polynomials for time-dependent
effects. Statistics in Medicine 2005;24:38713885.
[16] Remontet L, Bossard N, Belot A, Estève J, French network of cancer registries FRANCIM.
An overall strategy based on regression models to estimate relative survival and model the
effects of prognostic factors in cancer survival studies. Stat Med 2007;26:22142228.
114 REFERENCES
[17] Brenner H, Hakulinen T. Are patients diagnosed with breast cancer before age 50 years
ever cured? Journal of Clinical Oncology 2004;22:432438.
[18] Corazziari I, Quinn M, Capocaccia R. Standard cancer patient population for age standar-
dising survival ratios. Eur J Cancer 2004;40:23072316.
[19] Cronin KA, Feuer EJ. Cumulative cause-specific mortality for cancer patients in the pres-
ence of other causes: a crude analogue of relative survival. Statistics in Medicine 2000;
19:17291740.
[20] Lambert PC, Dickman PW, Nelson CP, Royston P. Estimating the crude probability of
death due to cancer and other causes using relative survival models. Stat Med 2010;29:885
895.

[21] Lambert PC, Dickman PW, Osterlund P, Andersson TML, Sankila R, Glimelius B. Tem-
poral trends in the proportion cured for cancer of the colon and rectum: a population-based
study using data from the Finnish cancer registry. International Journal of Cancer 2007;
121:20522059.
[22] Lambert PC, Thompson JR, Weston CL, Dickman PW. Estimating and modeling the cure
fraction in population-based cancer survival analysis. Biostatistics 2007;8:576594.
[23] Falcaro M, Nur U, Rachet B, Carpenter JR. Estimating excess hazard ratios and net
survival when covariate data are missing: strategies for multiple imputation. Epidemiology
2015;26:421428.
[24] Carpenter JR, Kenward MG. Multiple imputation and its application. Chichester: John
Wiley & Sons, 2013.

Labs

Uploaded by

Copyright:

Available Formats

Labs

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Labs

Uploaded by

Copyright:

Available Formats

Statistical methods for population-based cancer survival analysis

Computing notes and exercises

Paul W. Dickman1 , Paul C. Lambert1,2 , Sandra Eloranta1 ,

(1) Department of Medical Epidemiology and Biostatistics

(2) Department of Health Sciences

Cancer Research Centre, Cancer Institute of Iran,

1 Notes on survival analysis using Stata 4

2 Downloading user-written Stata commands and data files 5

203. Period estimation of relative survival . . . . . . . . . . . . . . . . . . . . . . . . . 72

1 Notes on survival analysis using Stata

A general introduction to Stata (stataintro.pdf) can be downloaded from:

. stset surv_mm, failure(status==1,2)

stset Declare data to be survival-time data

. sts graph, by(sex)

2 Downloading user-written Stata commands and data files

2.1 The quick and easy way

2.2 Downloading the course files

net install http://www.pauldickman.com/survival/course_files, all replace

2.3 Installing Stata user-written commands for relative survival

2.3.1 How can I check if these commands are already installed?

2.3.2 strs - estimating and modelling relative survival

. net install http://www.pauldickman.com/rsmodel/stata_colon/strs, all replace

2.3.3 stpm2 - flexible parametric models

ssc install stpm2

2.3.4 strsmix and strsnmix - cure models

2.3.5 Estimating probability of death in a competing risks framework

ssc install stcompet

The stpm2cif command is described in a Stata Journal article [5].

2.4 Relative survival in SAS and R

100. Hand calculation: Life table and Kaplan-Meier estimates of survival

Age Clinical dx date Surv. time

. use colon_sample, clear

. recode status (1=1) (nonmissing=0), gen(csr_fail)

. ltable surv_yy csr_fail

Alternatively, we could use

. ltable surv_mm csr_fail, interval(12)

. stset surv_mm, failure(status==1)

A listing of the Kaplan-Meier estimates is then obtained as follows

To graph the Kaplan-Meier estimates

. sts graph, atrisk

Titles and axis labels can also be specified.

. sts graph, risktable ///

102. Localised melanoma: Comparing actuarial and Kaplan-Meier approaches with

. use melanoma, clear

. stset surv_mm, failure(status==1)

. graph combine anydeathbyage cancerdeathbyage

104. Localised melanoma: Comparing estimates of cause-specific survival between

. use melanoma if stage == 1, clear

. use diet, clear

NOTE: Rates are calculated as events/person-time so when modelling rates we need

Compare your estimates with those you obtained in 110e.

111. Localised melanoma: model cause-specific mortality with Poisson regression

In this exercise we model, using Poisson regression, cause-specific mortality of patients

. use melanoma if stage==1, clear

(a) Plot Kaplan-Meier estimates of cause-specific survival as a function of calendar period

Is the interpretation similar? Do you think it is sufficient to classify follow-up time

. use diet, clear

.* Timescale: Attained age

. poisson chd hieng, e(y) irr