Interpretation of rank tests for kaplan meier

[Susan Maree Cotton]

Hi All,


I am analyzing some results for a psychotherapy RCT and have done survival
analysis with the survival being no relapse of symptoms.



I have the situation where the log-rank (Mantel Cox) p = .053, but the
Breslow p=.023, and the Tarone-Ware p=.026.



I have the following questions:

*       Does anyone have any references on the differences between these
three tests?
*       I know that in many studies, researchers will report all three
results to support differences in survival curve, but what do you do in the
situation that one test is not-significant the other two are significant.



Thanks in advance for your assistance



Regards,


Sue





Dr Sue Cotton
Senior Research Fellow (Statistician)

ORYGEN Research Centre

Department of Psychiatry

University of Melbourne

Locked Bag 10 (35 Poplar Road)

Parkville Victoria

Australia 3052

Phone: [phone]

Mobile: [phone]

[Reutter, Alex]

The short of it is that the log-rank method is best at detecting differences between the curves late in the time period of the study; the Breslow test is best at detecting early differences, and Tarone-Ware is an intermediate strategy.

Have a look at:

Hosmer, D. W., and S. Lemeshow. 1999. Applied Survival Analysis. New York: John Wiley and Sons.

Kleinbaum, D. G. 1996. Survival Analysis: A Self-Learning Text. New York: Springer-Verlag.

Norusis, M. SPSS Advanced Statistical Procedures Companion. Upper Saddle-River, N.J.: Prentice Hall, Inc..

These are taken from the "Recommended Readings" in the Kaplan-Meier "case studies" (Help > Case Studies).

Hosmer and Lemeshow give an example in which the log-rank and Breslow (which they call Wilcoxon) tests do not agree.

Cheers,
Alex


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Sue Cotton
Sent: Saturday, October 06, 2007 8:39 PM
To: [hidden email]
Subject: Interpretation of rank tests for kaplan meier

Hi All,


I am analyzing some results for a psychotherapy RCT and have done survival
analysis with the survival being no relapse of symptoms.


I have the situation where the log-rank (Mantel Cox) p = .053, but the
Breslow p=.023, and the Tarone-Ware p=.026.


I have the following questions:

*       Does anyone have any references on the differences between these
three tests?
*       I know that in many studies, researchers will report all three
results to support differences in survival curve, but what do you do in the
situation that one test is not-significant the other two are significant.


Thanks in advance for your assistance


Regards,
Sue





Dr Sue Cotton
Senior Research Fellow (Statistician)

ORYGEN Research Centre

Department of Psychiatry

University of Melbourne

Locked Bag 10 (35 Poplar Road)

Parkville Victoria

Australia 3052

Phone: [phone]

Mobile: [phone]

[Albert-Jan Roskam]

Hi listers,

Every now and then I make a mistake with the
following:

* sample data.
data list fixed / x 1-1 y 3-3 z 5-5.
begin data
1 1 1
1 1
1

end data.

* now consider the following command:.
if (max (x, y, z) ne 1) flag = 1.

Here I would find it logical that the last row (with
only NULL values) is evaluated as flag = 1. But it's
not! None of the records are evaluated as flag = 1. I
find that illogical, and therefore I make mistakes
with this every now and then.

* this gives the desired results.
recode x y z (sysmis = 0) (else = copy).
if (max (x, y, z) ne 1) flag2 = 1.

Could somebody explain me the logic behind this? Maybe
that way I won't make the same mistake again!

Thanks!

Albert-Jan

Cheers!
Albert-Jan

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Did you know that 87.166253% of all statistics claim a precision of results that is not justified by the method employed? [HELMUT RICHTER]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~


      ____________________________________________________________________________________
Fussy? Opinionated? Impossible to please? Perfect.  Join Yahoo!'s user panel and lay it on us. http://surveylink.yahoo.com/gmrs/yahoo_panel_invite.asp?a=7

[Peck, Jon]

Well, how would you know whether a missing value is equal to 1 or not?  It is just unknown.  The statistical procedures generally exclude sysmis values, so it makes sense for transformations to do likewise.  You will find that even the transformation
compute sm = $sysmis eq $sysmis
will produce $sysmis values for sm.

Dealing with three-valued logic, where the outcomes are true, false, and unknown, can certainly be confusing, but it reflects reality in many situations.

HTH,
Jon Peck

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Albert-jan Roskam
Sent: Monday, October 08, 2007 10:41 AM
To: [hidden email]
Subject: [SPSSX-L] MAX (Null, Zero, Value)

Hi listers,

Every now and then I make a mistake with the
following:

* sample data.
data list fixed / x 1-1 y 3-3 z 5-5.
begin data
1 1 1
1 1
1

end data.

* now consider the following command:.
if (max (x, y, z) ne 1) flag = 1.

Here I would find it logical that the last row (with
only NULL values) is evaluated as flag = 1. But it's
not! None of the records are evaluated as flag = 1. I
find that illogical, and therefore I make mistakes
with this every now and then.

* this gives the desired results.
recode x y z (sysmis = 0) (else = copy).
if (max (x, y, z) ne 1) flag2 = 1.

Could somebody explain me the logic behind this? Maybe
that way I won't make the same mistake again!

Thanks!

Albert-Jan

Cheers!
Albert-Jan

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Did you know that 87.166253% of all statistics claim a precision of results that is not justified by the method employed? [HELMUT RICHTER]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~


      ____________________________________________________________________________________
Fussy? Opinionated? Impossible to please? Perfect.  Join Yahoo!'s user panel and lay it on us. http://surveylink.yahoo.com/gmrs/yahoo_panel_invite.asp?a=7