Programming the statistical procedures from SAS

Very high hazard ratio from Proc Phreg

Reply
Contributor
Posts: 40

Very high hazard ratio from Proc Phreg

I ran proc lfietest to generate KM curves using the following code:

proc lifetest data= dat;

time Survival*censor(1) ;
strata group   /test=(logrank);
run;

 

Since I wanted the hazard ratios for this, I ran the following Proc Phreg code:

proc phreg data=dat;
class group;
model Survival*censor(1) =group ;
run;

 

The HR came out to be very high -a seven digit number (3797239) and not significant - pvalue=0.99. Can someone help me understand this?

 

Thanks so much!

Super User
Posts: 18,569

Re: Very high hazard ratio from Proc Phreg

Can you post a PROC means on your survival time BY group and by censor. I suspect the proportionality assumption is violated for one thing, I think the ASSESS statement can be used to check this assumption. 

 

If you could post the results that would be helpful. 

Contributor
Posts: 40

Re: Very high hazard ratio from Proc Phreg

Thank you so much for helping with this. Any insight you could give with this is highly appreciated since I only have a couple of hours to deliver these HRs

 

Here are the Proc means: (I will be checking assess shortly)

The MEANS Procedure    
Group=No      
Analysis Variable : Survival Time [Days]
N Mean Std Dev Minimum Maximum
107 756.79 355.51 9 1246
Group=Yes      
Analysis Variable : Survival Time [Days]
N Mean Std Dev Minimum Maximum
35 868.8 242.74 250 1141
         
         
The MEANS Procedure    
Censoring variable =No    
Analysis Variable : Survival Time [Days]
N Mean Std Dev Minimum Maximum
16 406.88 158.85 196 701
Censoring variable =Yes    
Analysis Variable : Survival Time [Days]
N Mean Std Dev Minimum Maximum
126 832.33 319.79 9 1246
Super User
Posts: 18,569

Re: Very high hazard ratio from Proc Phreg

You only have 16 uncensored events? That's a small rate which is contributing to the issue. Especially when you then break it down into groups as well. 

Contributor
Posts: 40

Re: Very high hazard ratio from Proc Phreg

Thanks

Attached is also the KM plot as well... 

What would be a good clinical or statistical reason for not reporting the hazard ratio?


KM plot.PNG
Super User
Posts: 18,569

Re: Very high hazard ratio from Proc Phreg

It appears the Proportionality assumption required for a Cox regression is violated. This is compounded by the fact that you have very low event rates and really not a lot of observations. Therefore estimates from this data would have a large degree of uncertainty around  them. 

Contributor
Posts: 40

Re: Very high hazard ratio from Proc Phreg

Thank you so much for the help. I will update later on running assess statement

Super Contributor
Posts: 287

Re: Very high hazard ratio from Proc Phreg

Such big estimates only occur when the model doesnt converge. I think that there is no uncensored events in one of the two groups. It can even occur if there is uncensored events in both groups, but only if all uncensored events in one group comes when all persons in the other group are no longer in risk.

Try make the proc means again with censoring and group at the same time in the by statement.

I dont think it has to do with the the assumption of the proportionality statement.

Contributor
Posts: 40

Re: Very high hazard ratio from Proc Phreg

Thanks for your comments. Here are the mean measures:

The MEANS Procedure
Censoring variable=No Group=No
Survival Time [Days]
N Mean Std Dev Minimum Maximum
11 396.73 172.71 196 701
Censoring variable =No Group=Yes
Survival Time [Days]
N Mean Std Dev Minimum Maximum
5 429.2 138.35 250 578
Censoring variable =Yes Group=No
Survival Time [Days]
N Mean Std Dev Minimum Maximum
96 798.04 348.07 9 1246
Censoring variable =Yes Group=Yes
Survival Time [Days]
N Mean Std Dev Minimum Maximum
30 942.07 166.05 294 1141

 

As for the convergence, I don't get any errors/warnings. It says "convergence criteria satisfied"...

Super Contributor
Posts: 287

Re: Very high hazard ratio from Proc Phreg

It seems strange that the hazardratio should be that extreme. Can you attach the dataset? I would like to take a look on it.

Ask a Question
Discussion stats
  • 9 replies
  • 194 views
  • 0 likes
  • 3 in conversation