Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
st: RE: Proportional hazard assumption for interaction term
From
"Kieran McCaul" <[email protected]>
To
<[email protected]>
Subject
st: RE: Proportional hazard assumption for interaction term
Date
Tue, 25 May 2010 08:25:23 +0800
....
OK, I'll have a go at this.
(1) In your model, you've fitted thincat99, which as you stated, indicates missing thincat values. You've also fitted stage9 which I assume means missing stage.
You can't treat missing values like this in a model because they are not a separate category, they are some unknown mixture of the categories. So you either drop them from the model or use imputation to impute the missing data. You need to deal with this first before you start worrying about proportionality assumptions.
(2) Some of the variables with significant p-values from the phtest also have very small rho values. This is the case for thincat5, which is the effect of thincat in yearcat1. I'm guessing you have a large dataset with a large number of events observed, so the phtest will have power to detect small non-proportionality effects. You need to plot the Schoenfeld residuals over time, overlayed with a lowess smoother, to see what is going on here. This is explained in the Stata manual.
(3) The lack of proportionality with the stage variable is not surprising if, as I suspect, this is a cancer dataset and the outcome is death or recurrence or both. Stage is a fairly crude classification and there will still be a fair degree of heterogeneity of risk within each of these categories. So, early in the course of follow-up, those most at risk in each stage category fail first, and over time the heterogeneity in risk within each stage category is reduced. This happens relatively quickly, with a year or so after diagnosis, and is the source of the lack of proportionality. If you graph the hazard curves you'll see what I mean. How you deal with this depends on what you want from the model. If you aren't interested in the stage-specific hazard ratios then you could simply stratify on stage.
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sanam P
Sent: Monday, 24 May 2010 8:23 PM
To: [email protected]
Subject: st: Proportional hazard assumption for interaction term
Dear Statalist
I was wondering if in the Cox model there is interaction between two categorical variables (here thincat and yearcat), when checking the proportional hazard assumption using :
"estat phtest,rank detail" after the model, for the interaction term and the variables involved in interaction to be proportional, should all the terms involved in the interaction be unsignificant? if only one of the main effects at reference value of the other variable involved in the interaction term is significant and all other terms involved in the interaction are not significant, does it still mean that variable is not proportional? is the interaction proportional?
(thincat=99 is a category with missing thincat values)
Thank you very much in advance,
Sanam
Time: Rank(t)
----------------------------------------------------------------
| rho chi2 df Prob>chi2
------------+---------------------------------------------------
_Ithincat~2| 0.00850 0.38 1 0.5365
_Ithincat~3| 0.00776 0.32 1 0.5725
_Ithincat~4| -0.01565 1.30 1 0.2546
_Ithincat~5| -0.04559 11.07 1 0.0009
_Ithinca~99| -0.04752 11.73 1 0.0006
_Iyearcat_2 | -0.00202 0.02 1 0.8835
_Iyearcat_3 | 0.00318 0.05 1 0.8179
_Iyearcat_4 | -0.00582 0.18 1 0.6727
_IthiXye~2_2| -0.00935 0.46 1 0.4964
_IthiXye~2_3| -0.00543 0.16 1 0.6928
_IthiXye~2_4| 0.00142 0.01 1 0.9179
_IthiXye~3_2| -0.00478 0.12 1 0.7281
_IthiXye~3_3| -0.00690 0.25 1 0.6158
_IthiXye~3_4| -0.00593 0.19 1 0.6659
_IthiXye~4_2| -0.00334 0.06 1 0.8081
_IthiXye~4_3| -0.00142 0.01 1 0.9177
_IthiXye~4_4| 0.00910 0.44 1 0.5083
_IthiXye~5_2| 0.00079 0.00 1 0.9543
_IthiXye~5_3| 0.01261 0.84 1 0.3594
_IthiXye~5_4| 0.01758 1.63 1 0.2013
_IthiXye~9_2| 0.00488 0.13 1 0.7229
_IthiXye~9_3| 0.01125 0.67 1 0.4141
_IthiXye~9_4| 0.02605 3.58 1 0.0584
_Istage_2 | -0.07686 33.18 1 0.0000
_Istage_3 | -0.12396 82.46 1 0.0000
_Istage_99 | -0.01765 1.65 1 0.1995
_Ihepcat_2 | -0.02429 3.22 1 0.0727
_Ihepcat_3 | 0.02983 4.73 1 0.0296
_Ihepcat_4 | -0.02254 2.58 1 0.1085
_Ihepcat_5 | 0.03376 6.11 1 0.0135
_Ihepcat_6 | 0.01624 1.43 1 0.2324
_Ihepcat_7 | 0.00075 0.00 1 0.9562
_Isitecat_2 | -0.02876 4.56 1 0.0327
_Isitecat_3 | -0.02031 2.24 1 0.1345
_Isitecat_4 | -0.01405 1.09 1 0.2973
_Isitecat_5 | -0.11526 65.59 1 0.0000
_Isexcat_1 | 0.01412 1.08 1 0.2979
_Iseason_2 | 0.01397 1.03 1 0.3092
_Iseason_3 | 0.00282 0.04 1 0.8368
_Iseason_4 | 0.02057 2.24 1 0.1344
agecatf | -0.00911 0.49 1 0.4843
------------+---------------------------------------------------
global test | 976.68 41 0.0000
----------------------------------------------------------------
*
* 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/
*
* 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/