.-
help for ^stcascoh^ [ec] Nov 21, 2008
.-
Create a dataset suitable for case-cohort analysis
--------------------------------------------------
^stcascoh^ [varlist] [^if^ exp] [^in^ range] ^, a^lpha^(^#^)^ [ ^gr^oup^(^varnames^)^
^gen^erate^(^varlist^)^ ^ep^s^(^#^)^ ^se^ed(#) ^nosh^ow ]
^stcascoh^ is for use with survival-time data; see help @st@. You must ^stset^ data with
an ^id()^ variable before using this command; see help @stset@.
Description
-----------
^stcascoh^ is used to create the appropriate dataset for a case-cohort analysis by
sampling the cohort at the time of entry (subcohort) and including all failures, whether they occur in
the random sample or not (non subcohort cases). To this aim, ^stcascoh^ expands the observations who fail in two
parts: (1) time interval (t0,t-eps] and (2) time interval (t-eps,t].
The following variables are added to the dataset:
^ _subco^ coded 0 for subcohort member with no failure
1 for subcohort members who failed
2 for non subcohort cases
^ _wBarlow^ log-weights of records as in Barlow scheme
^ _wSelPre^ log-weights of records as in Self and Prentice scheme.
The names of the new variables can be changed by specifying the ^gen()^ option.
varlist defines the variables to be retained in the final dataset. If varlist is not specified, all variables
are carried over into the resulting dataset.
^stcascoh^ shows two summary tables: the first one describes the subcohort membership
vs. failure in the cohort; the second displays the risk sets with three controls or less
to check whether the subcohort becomes small due to many failures or censorings.
In the new dataset, non-subcohort cases cannot rely on the original ^stset^ declaration.
At the end of the module, ^stset^ is invoked to fix entry and exit time to the present
_t0 and _t variables.
Randomness in the sampling is obtained using Stata uniform() function. Seed can be specified
by a ^seed()^ option. Observations not meeting ^if^ and ^in^ criteria are dropped even if
they fail.
The resulting dataset can be analyzed using ^stselpre^ or ^stcox^.
^stselpre^ fits proportional hazards model according with Prentice and Self-Prentice methods.
Self-Prentice model based variance is estimated.
When using ^stcox^, ^robust^ option is needed to estimate the approximate variance as proposed by Lin
and Ying and Barlow. In this case, analysis can be performed using three methods:
1- Prentice: ^stcox varlist, robust^
2- Self and Prentice: ^stcox varlist, offset(_wSelPre) robust^
3- Barlow: ^stcox varlist, offset(_wBarlow) robust^
Options
-------
^alpha(^#^)^ specifies the sampling fraction. Sampling fraction can be expressed as real or
integer.
^group(^varlist^)^ specifies that the ^alpha^ sample is to be drawn within each set of values
of varlist, thus maintaining the proportion of each group.
^generate(^varlist^)^ specifies variable names for three generated variables.
^eps(^#^)^ specifies a typically small number so that a case that is in the risk at time
set at t is represented in the expanded data by an "infinitesimal" episode (t-eps,t].
^eps^ should be set to a number that is small compared to the measurement unit of time.
^eps^ defaults to 1E-3.
^seed(^#^)^ specifies seed for random sampling.
^noshow^ prevents ^stcascoh^ from showing the names of the key ^st^ variables.
Tip: if the cohort has already been sampled
-------------------------------------------
^stcascoh^ can be used to prepare the dataset for case-cohort analysis when the cohort
has been previously sampled. The steps are as follows:
1) Divide the data set in two files: the first one for subcohort observations, the second for additional cases.
2) Appropriately -stset- the new files.
3) Process the subcohort file with -stcascoh- setting alpha just below 1 (i.e. .999) and save this file.
Do some assert to verify that _subco _wSelpre are coded as you expect.
4) Process the additional cases file with -stcascoh- setting alpha = 0 and save this second file.
Also for this file some -assert- is appropriate to verify that _subco _wSelpre are coded as you expect.
5) Append the two files you generated.
6) Adapt the Barlow weights. Suppose you sampled 1000 out of 16000 subjects. You should type
^replace _wBarlow=ln(16000/1000) if _wBarlow>0^
Examples
--------
. ^stcascoh, alpha(20)^
. ^stcascoh afe yfe ln_exp, alpha(0.3) gen(mycohort) group(race) seed(987654321)^
Also see
--------
Manual: ^[R] stcox [R] sttocc^
On-line: help for @stcox@ @stselpre@ @sttocc@
Reference
---------
Barlow WE, Ichicawa L, Rosner D, and Izumi S: Analysis of Case-Cohort Designs.
Journal Clinical Epidemiology 1999; 52: 1165-1172.
Author
------
Enzo Coviello
Unita' di Epidemiologia e Statistica Az. USL BA/1
70053 Andria (Bari)
Italy
enzo.coviello@@alice.it