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: Re: statistical test and sensitivity analysis for matched pairs with censoring

From   "Shoryoku Hino" <[email protected]>
To   <[email protected]>
Subject   RE: st: Re: statistical test and sensitivity analysis for matched pairs with censoring
Date   Sat, 1 Jan 2011 15:34:05 +0900

Joseph and Steve,

Your suggestions are great help. Related to the tests for matched pair, I am
going to make it with -sts test-or -stcox- using pair strata. I had actually
considered IPTW methods as one of the other options. 
But as Joseph said, is there any tailored method of -rbounds- or -mhbounds-
to survival censored data as outcome?

-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Steven Samuels
Sent: Wednesday, December 29, 2010 1:29 AM
To: [email protected]
Subject: Re: st: Re: statistical test and sensitivity analysis for matched
pairs with censoring

Shoryoku Hino

Also, try -stcox- with pair-strata.  See page 330 of  Prentice and
Kalbfleisch, 2002, The Statistical Analysis of Failure Time Data, 2nd Ed,
Wiley, NY.  Adding the -frailty- option is not necessary, but adding
-vce(cluster pair_id)- would be worthwhile. (The vce(cluster) option is not
available in -sts test-.)

You can also do IPTW (ATT etc.) with propensity scores in -stcox- by   
setting up the data with probability weights via -svyset- (as well as -
stset-). In that case it is immaterial whether you specify the pair id as a
PSU in -svyset- or in the -vce(cluster pair_id)- option.

Compared to what has been proposed in the literature, the choices in Stata
are not great.  Many references can be found in Section 5 of Shih, Joanna
and Michael Fay (2003), 'Chapter 8: A Class of Permutation Tests for Some
Two-Sample Survival Data Problems,', in Geller, Nancy L. (ed.), Advances in
Clinical Trial Biostatistics, CRC Press).

Note that 1:1 matching is not an optimal design, especially for survival
data. N:M  matching would be more informative in general, but with survival
data, there's another reason to do it: a stratum with all members censored
contributes nothing to the analysis, and the risk of this is greatest when


Steven J. Samuels
[email protected]
18 Cantine's Island
Saugerties NY 12477
Voice: 845-246-0774
Fax:    206-202-4783

On Dec 28, 2010, at 9:27 AM, Joseph Coveney wrote:

Shoryoku Hino wrote:

Thank you for your reply and kind advice. I should have made my question
All I want to know is the test for survival data in case of paired sample.

For example, if it were not survival data, we would use signed rank test for
continuous variable in pared sample instead of unpaired t-test or rank sum
test. I would use McNemar test for proportion instead of chi square test.
Rosenbaum's sensitivity test could be applicable in these cases.

In my case, survival data, I think there might be a better, more powerful
test for paired sample than log-rank test or Wilcoxon test.
I would like you to give me any suggestion.


There is a -strata()- option to -sts test- that would allow you to stratify
on pairs.  I don't know whether this achieves the same result as does what
your reference ("RF.Woolson") is talking about.

Have you considered modeling the survival time with -stcox- or -streg- in
lieu of paired-sample testing?  The model would include some or all of the
confounder covariates in addition to the propensity score or an indicator
variable for matched pair.  See, for example, A. Gelman & J. Hill, _Data
Analysis Using Regression and Multilevel/Hierarchical Models_ (New York:
Cambridge Univ. Press, 2007), pp. 206-12; J. Hill, Discussion of research
using propensity-score matching: Comments on 'A critical appraisal of
propensity-score matching in the medical literature between 1996 and 2003'
by Peter Austin, Statistics in Medicine. _Statistics in Medicine_
2008 (
and references cited in them.

Another alternative to a simple paired-sample test might be using the scores
in a weighted regression model, which would also include some or all of the
confounder covariates; see, for example, A. Nichols, Causal inference with
observational data. _Stata Journal_ 7:507-41, 2007; A. Nichols, Erratum and
discussion of propensity-score reweighting. _Stata Journal_ 8:532-39, 2008.

For Rosenbaum-bounds and related analysis, there are user-written Stata
commands -rbounds-, -sensatt- and -mhbounds-.  The last has an associated
article (S. O. Becher & M. Caliendo, Sensitivity analysis for average
treatment effects. _Stata Journal_ 7:71-83, 2007) that refers to what had
been implemented in Stata up to then.  Again, I don't know of anything
specifically tailored to censored survival time as the outcome.

Joseph Coveney

*   For searches and help try:

*   For searches and help try:

*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index