Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

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

Re: st: intcens: how to estimate mean and variance after intcens

From   Yoann Madec <>
To   Steve Samuels <>,
Subject   Re: st: intcens: how to estimate mean and variance after intcens
Date   Fri, 26 Oct 2012 10:03:20 +0200

Thanks Steve for your messages.

However, I may have to provide some clarifications.

I used :
stset date_deb_periode1, scale(365.25) origin(seroco_d)

1- date_deb_periode1 and seroco_d are dates, and scale(365.25) enables results to be expressed in years. But I could easily omit the scale option.

2- All patients experienced the event, this is why failure() does not appear in the stset command.

3- The event I am interested in occurs between seroco_d and date_deb_periode1, thus the use of interval-censored methods.

In the help document regarding stpm, it is said:
left(leftvar) specifies that some or all of the survival-time observations are interval-censored. The rules for specifying values of leftvar and their meanings in terms of interval censoring are as follows:
        Value of leftvar   _d   Meaning
        . or _t             0   Right censored at _t
        . or _t             1   Event at _t
        0                   0   Right censored at _t
        0                   1   Interval censored, event in (0,_t]
        <_t                 0   Late entry at leftvar, right censored at _t
        <_t                 1   Interval censored, event in [leftvar,_t]

In order to test whether stpm was working ok with simple data, I first omitted the left command (as if all events occured at time _t).
That worked ok.

Then I used the left(seroco_d) option. I know that for every patient seroco_d is strictly before date_deb_periode1.
xi: stpm i.gender if num_v==1, stpmdf(6) scale(hazard) left(seroco_d)
i.gender          _Igender_1-2        (naturally coded; _Igender_1 omitted)
seroco_d>_t in some observations

While it is not true that seroco_d is >date_deb_periode1.

I thought that maybe the fact that within the stset, I also used seroco_d in the origin() option. Therfore, I created artificially a new variable simply the mid-point between seroco_d and date_deb_periode1, but still have the same problem and the same error message.

I hope someone can help.


Le 26/10/2012 03:20, Steve Samuels a écrit :

If the origin of observation is truly seroco_d, your date of left
censoring, then you do not have left-censored or interval-censored data.
You have ordinary right-censored data.

In that case, you can use -stpm2- (from SSC), which does not accept
interval-censored data, but otherwise has some advantages over -stpm-.

I suggest that study the section "Two concepts of time" in of the Manual
entry for -stset- .



On Oct 25, 2012, at 4:43 AM, Yoann Madec wrote:

Thanks steve for your comments, and sorry not to have mention the source of the intcens command.

After reading the help page for stpm, I still do not manage to make it work with interval-censored data.

As a test, I have written:

	gen date_left_censoring=seroco_d
	stset date_deb_periode1, scale(365.25) origin(seroco_d)


In this case, I should have no interval-censored data, but strictly right-censored data.
However, here is what STATA states:

.         xi: stpm i.gender if num_v==1, stpmdf(6) scale(hazard) left(date_left_censoring)
i.gender          _Igender_1-2        (naturally coded; _Igender_1 omitted)
date_left_censoring>_t in some observations

I have been trying many things without success.

I hope someone can help.

Best regards,

Le 12/10/2012 19:48, Steve Samuels a écrit :

Yoann, The FAQ ask that you state the source of unofficial commands.
-intcens- was written by Jamie Griffin and is available from SSC.

The usual sample descriptive statistics cannot be calculated for
interval-censored data.

One approach is to apply Patrick Royston's command -stpm-, also at SSC,
which fits flexible distributions. You can estimate survival curves and
percentiles of the unconditional as well as covariate-conditional,
distributions. You won't get standard errors for the percentiles, but
you could -bootstrap- these. Means and SDs can be estimated with a lot
more work, but I don't think these are useful descriptive stats for
most survival data problems.

In fact, I recommend -stpm-, not -intcens- for your main analysis. Some
reasons: 1) Both fit parametric models, but -stpm- adapts to the shape
of the distribution, saving you the need to select a "best" theoretical
distribution. 2) -stpm- has excellent postestimtion options; -intcens-
has none. (You can estimate survival curves&   statistics starting with
the supplied e(b) matrix, but you must do it by hand.) 3) -stpm- allows
coefficients to vary with time (i.e. time-predictor interactions,);
-intcens- does not.


On Oct 11, 2012, at 12:35 PM, Yoann Madec wrote:

Dear Stata users,

In order to describe the time to an event I used the command intcens. Indeed, for all my subjects, I know that the event took place within a time-interval, but do not have thje exact date.

Using intcens, I vcan test whether some factors influence this time to event.

However, I would like to summarize the time to event and provide a confidence interval for this time.
I have not been able to fin how to estimate a mean and variance after intcens.

i hope that someone will be able to help.

Best regards,


*   For searches and help try:

*   For searches and help try:

*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index