Date:         Fri, 21 Jan 2005 10:17:55 +1030
Reply-To:     Kylie Lange <kylie.lange@flinders.edu.au>
Sender:       "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From:         Kylie Lange <kylie.lange@flinders.edu.au>
Subject:      Re: survival analysis with left censored data
Comments: To: Pablo Mora Olmedo <pmora@puc.cl>
In-Reply-To:  <20050120230032.84038.qmail@smtp.puc.cl>
Content-Type: text/plain; charset="us-ascii"; format=flowed

Hi Pablo,

As far as I am aware methods for arbitrarily censored data (ie, data that
contains a mixture of different censoring types) are restricted to full
parametric models. As you say, Cox regression (semi-parametric) is
applicable for right-censored data only. Parametric models (Weibull,
log-linear, log-normal, gamma etc) are not implemented in SPSS however
there are procedures in SAS, Minitab and S-Plus among others.

I do recall reading somewhere that the non-parametric method called
Turnbull's method (analogous to Kaplan-Meier but for left and right
censored data) can be modified to handle interval censored data, but I
don't remember any of the details.

Some of my favourite survival references that include parametric models:

D. Collett. 1994. Modelling Survival Data in Medical Research
J.P. Klein & M.L. 1997. Moeschberger. Survival Analysis: Techniques for
Censored and Truncated Data.
D.W. Hosmer & S. Lemeshow. 1999. Regression Modeling of Time to Event Data.

 From memory some of these references mention extensions to Cox regression
that can handle censoring schemes other than right-censoring, but when I
looked into it about a year ago none of those methods had reached the stage
of being commonly used in the applied literature or implemented in standard
software packages.

Hope this helps.

Cheers,
Kylie.


At 11:00 PM 20/01/2005 +0000, Pablo Mora Olmedo wrote:
>Right to the problem, I need to model time-to-event in a sample with a good
>number of left- and some interval-censored cases.  My impression is that I
>can't use Cox or proportional hazards regression? Am I right? If that's the
>case can anyone give me some references I could use to find out what type of
>analysis is the most appropriate?
>Thanks in advance
>Pablo
>BTW, Thanks to all the people who kindly answered my email regarding survey
>software.