[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: AW: RE: Problem looping over spells for an individual

From   Ilona Carneiro <>
Subject   Re: st: AW: RE: Problem looping over spells for an individual
Date   Fri, 27 Feb 2009 19:09:14 +0100

Thanks for these suggestions, Nick. Some of them seem to require "tsset"ing the data. However, I am trying to adapt the code from panelthin.ado, but can't quite get it to work for me. I've been battling with this for a couple of days now and I can't seem to find a way to loop over consecutive observations for an individual. I see that in panelthin, it automatically works separately on each panel because it is tsset, but how can I do this for survival time "stset" data?

I've written a little sub-programme to see if I can get this to work. Generation of a local for _N now works and does give the correct count of observations per individual. However, my generation of the tempvar `t' to define the sequential observations (_n) for each individual doesn't work. Any suggestions?

		capture program drop temp		
		program define temp, byable(recall, noheader) sortpreserve
		marksample touse
		count if `touse'
		if r(N) == 0 error 2000

		tempvar T t
		sort `_byvars' start
		by `_byvars': gen `t' = _n * `touse'
		sort `_byvars' start
		by `_byvars': gen `T' = _N * `touse'
		sort `_byvars' start
		sum `T', meanonly
		local tmax = r(max)
		drop `T'

		replace lagend = (end + 19 + 1) if (anmal > 0 & anmal < .)
		sort pin start
		forvalues i = 1(1)`tmax' {
			drop if end < lagend[`i'-1] & lagend[`i'-1] < . & `t'==`i' & `i'!=1
replace lagend = (end + 21 + 1) if (mal0 > 0 & mal0 < .) & lagend==. & `i'<`tmax'
		bysort id: temp



On 26 Feb 2009, at 13:20, Nick Cox wrote:

Thanks for this, which is good news for me because it explains why the code I was seeing looked as it did.

In terms of moving forward, I have a few vague suggestions.

0. Spells. See the suggestions on reading and software in the thread started by Jakob Petersen yesterday.

< >

1. One is more of style or taste than technique. I prefer to think in terms of tagging observations I want to keep or work on with 1 and those I don't with 0. Then you can do almost anything later

	... if tag


	... if !tag

as the case may be.

An advantage of that style: it is reversible, both within an algorithm and generally. (If you really want to -drop- observations, -drop- them in one go when the selection is final.)

2. One strategy might be

loop over individuals {
-expand- each individual to a block of observations with one observation per day
	<magic bit>
	reduce each individual back again

3. This problem reminds me loosely of one tackled with -panelthin- on SSC. The code for that may suggest some technique.


Ilona Carneiro

Many thanks to Nick & Martin for pointing out my error using "if" -
you are correct and that's why it wasn't working. However, I'm still
unable to do what I wanted to. Apologies for posting code which I
tried to simplify, but just made incomprehensible! The snippet was
part of a much larger programme in which the other local macros are
all defined.

I'll try to clarify. Here is an example of the problem I have. These
are consecutive periods of observations for an individual - the end
denoted by a clinic visit which may or may not be defined as a case
(depending on diagnostic result), or by exit from the study.

id		start		end		case	tx
1		10		20		1		1
1		20		35		1		0
1		35		50		1		0
1		50		100		.		.

I need to exclude 19 days at risk if the patient received treatment
(tx==1) as this is considered to be prophylaxis, and to avoid counting
the same episode (case==1) twice I  also exclude 19 days at risk after
a case is diagnosed. However, as the latter is only to prevent double-
counting it is not necessary if the case has already been disqualified.

What I need to get is the following:

id		start		end		case	tx
1		10		20		1		1
1		40		50		1		0
1		50		100		.		.

I originally coded the following VERY crudely:

/* To calculate the gaps  */
sort id start
by id: gen lagend = end + lag if (tx > 0 & tx < .) | (case > 0 & case
< .) & _n!=_N

/* To drop periods of time that are disqualified - repeated 3 times as
there may be up to 3 consecutively - to be generalisable, it could be
more */
sort id start
by id: drop if lagend[_n-1] > end & lagend[_n-1] < . & _n!=1
sort id start
by id: drop if lagend[_n-1] > end & lagend[_n-1] < . & _n!=1
sort id start
by id: drop if lagend[_n-1] > end & lagend[_n-1] < . & _n!=1
sort id start
by id: drop if lagstart > end & lagstart < . & _n!=1

/* To update the start date */
sort id start
by id: replace start = lagend[_n-1] if lagend[_n-1] < . & _n!=1	
sort id start
by id: drop if (end < start | start[_n-1] > end) & end < . & start < .
& _n!=1

This works fine for adding a gap after each treatment, as I need to do
this even if the observation period is dropped from the time at risk.
The code gave the following result, as both the 2nd & 3rd episodes
were disqualified, instead of just the 2nd:

id		start		end		case	tx
1		10		20		1		1
1		55		100		.		.

I realise that I need to evaluate the generation of the gap after
cases separately for each observation period, incase the observation
is dropped. But can't seem to find a way to do this. I hope this is a
clearer explanation of the problem.

On another point, I subsequently use stgen gap =  gaplen() to
calculate how much time to exclude from the time at risk. Stata
appears to count one more than just the actual gap, i.e. it will give
me a gap of 20 days between an observation ending with day 20, and a
subsequent observation starting at day 40, when the actual time
excluded in-between is 19 days. I'm just subtracting 1 from the
calculation at present, but is there a reason for this?


On 25 Feb 2009, at 18:27, Martin Weiss wrote:


I was desperate to find an SJ tip for Ilona on the difference
between "if"
and "if"; turns out it is an FAQ:


-----Ursprüngliche Nachricht-----
[] Im Auftrag von Nick Cox
Gesendet: Mittwoch, 25. Februar 2009 18:22
Betreff: st: RE: Problem looping over spells for an individual

Unless you are working under the aegis of -by:- _N will always be
interpreted as the total number of observations. This code doesn't
satisfy that.

I echo Martin Weiss in suspecting that your -if `touse'- is a bug. You
are almost certainly confusing the two flavours of -if-.

Otherwise, your code still looks very confused and based on a
variety of
misunderstandings. Apart from `touse', which is defined by -
all of the local macros you refer to will be treated as empty strings,
as none has been defined earlier in the program. I am surprised to
that it is running at all.

It does not look as if you need a program anyway. My impression is
all you need is to use -by:- but I don't understand your problem well
enough to suggest better code. Someone else may be able to give better
help. If not, rather than a lengthy word description, you should
give an example of your data with the intended result.


Ilona Carneiro

I am trying to write a programme that will run a command sequentially
for observations of an individual. For each individual I have multiple
spells and multiple failures. However, the twist is that I also need
to exclude a period of time at risk after each treatment (prophylaxis)
and after each failure (to prevent double-counting of failures that
may actually be the same episode). I managed to do this without any
problem for the treatment, but if an episode is disqualified (by a
prior treatment or episode) I don't want it to disqualify a subsequent
episode. Therefore I need to run the code sequentially for each spell
of an individual, but using the marksample touse code to run it "by"
individual doesn't seem to be working - the "forvalues" seems to
always interpret _N as the last observation in the whole dataset, not
the last observation for each individual.

I have the following code:

		program define byid, byable(recall, noheader)
		marksample touse
		sort `id' `start'
		if `touse' {
		forvalues i = 1(1)`=_N' {
		replace lagend = (`end' + `lag') if ((`tx' > 0 & `tx' <
.) | (`case'
0 & `case' < .))
		drop if lagend[`i'-1]>`end' & `id'[`i'-1]==`id'
		gen lagend=. 	
		qui by id: byid

but I get the error:
2nd by group not found

And the programme isn't doing what I need it to.

*   For searches and help try:

*   For searches and help try:

*   For searches and help try:

*   For searches and help try:

*   For searches and help try:

© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index