Re: st: Interval censoring using intcens

[Steve Samuels <sjsamuels@gmail.com>]

> On Jul 28, 2012, at 4:29 AM, Patrick Munywoki wrote:
> 
> Hi,
> I have been attempting to analyse interval censored time-to-event data
> with 'intcens' ado (Griffin et al 2006). My data arise from a
> longitudinal household-based study with nasal swab collections
> twice-a-week for a duration of 26 weeks regardless of their any
> symptoms. I want to be able to estimate the duration of infectious
> period for one of the viruses we detected. I have reduced the data  to
> one observation per infection episode in order to use the 'intcens'
> command with t0 being the date last positive sample while t1 is the
> date of the next negative sample. I hope this data conversion to
> single observation per infection episode data is alright?

It looks OK to me. Note that a reference to Griffin et al. (2006) is not
very useful. Statalist FAQ ask you to give bibliography style references
and also provide the source for all non-official
commands (and that includes -stpm-) below.  It's sufficient to say
"from SSC" if that is where you got the command, but not all commands
are at SSC, so you have to give the link for those.

> 
> My questions?
> 1. How do i interpret the coefficient given in the results below?
> 
The exponential model assumes that there is a constant hazard of
failure in males and a different constant hazard in females.  
The coefficient is the log  of the ratio of male/female hazards. (The
"eform" option will give you results in terms of the ratios.)
These are pretty elementary facts in parametric survival analysis. If they
are not familiar to you, then I suggest that you study a good text
before proceeding.


> intcens t0 t1 male, dist(exp) time nolog
> 
> stata output
> Exponential distribution -      log acceleration factors
> 
> Uncensored               0
> Right-censored           0
> Left-censored            0
> Interval-censored      188
> 
>        Number of obs   =       188
>        Wald chi2(1)    =       0.00
> Log likelihood =  -1796.982     Prob > chi2     =       0.9990
> 
> 
> Coef.   Std. Err.      z    P>z     [95% Conf.  Interval]
> 
> male   -.0001871        .1470683    -0.00   0.999    -.2884356  .2880615
> _cons    9.817517       .2234524    43.94   0.000     9.379558  10.25548
> 
> Note the actual interval between the dates t0 and t1 is on average(sd)
> 3.6 (0.98) days; median(IQR) 4 (3-4) days; and range 2-7 days.
> 
> 
The standard errors are not valid, because they assume that observations are independent. They don't account for correlation of observations in the same individual or household. 

Add a cluster() option to identify the individual. If you have data on >1 person in 
a household, then let household ID  define the cluster.  If the presence of
others in the HH with with the infection can affect duration, then you should add 
indicators of this (e.g. presence of another infected person, number of infected people) to the individual data line. 

> 2. Whenever i try using any other distribution this error message pops
> up. What could be the problem here?
> intcens t0 t1 male, dist(weib) time nolog
> initial values not feasible
> r(1400);

 The -help- for -intcens- links to the help for -maximize-. 
Try adding the "difficult" option.  As the exponential is
unrealistic , you will want a more complex model. 


> 
> 3. Is there an alternative method to the interval censoring which
> allows me to use the multiple records per person accounting for the
> interval censoring. I have tried stpm but not sure whether it allows
> for this.
> 
-stpm- is a good choice and also has a cluster() option.


Steve
sjsamuels@gmail.com
 



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/