Discussion: View Thread

Splitting the survival curve

  • 1.  Splitting the survival curve

    Posted 09-18-2015 16:58

    Hello.

    Does anyone have any experience (or references to share) on how to analyze survival data which shows separation after a certain time point.  I have a dataset which shows survival curves crossing at roughly the mid-time point (log rank and Wilcoxon-Gehan tests are all non-significant), and wonder if I can analyze the data to help support a conclusion that once survival is reached at the mid time-point, there is a benefit with one product over another?  

    Would it be acceptable to partition the time axis somehow and re-run the non-parametric tests?

     

    Much thanks in advance.

    Shelley



    ------------------------------
    Shelley-Ann Walters
    3M
    ------------------------------



  • 2.  RE: Splitting the survival curve

    Posted 09-18-2015 17:07

    Shealley-Ann:

    Unless you prespecified the time point in advance, I wouldn't advise doing such data derived inference.  Have you looked at any of the non-directional tests of the survival curves, like a Cramer-Von Mises test? 


    ------------------------------
    Roy Tamura
    Associate Professor
    University of South Florida
    ------------------------------




  • 3.  RE: Splitting the survival curve

    Posted 09-19-2015 15:49

    Agree that unless threshold is defined in advance the finding needs to be verified. But a finding of significant non proportionality at least is evidence that relative group differences are not constant over time.
    **********************************
       






  • 4.  RE: Splitting the survival curve

    Posted 09-19-2015 08:12

    I've never done this, but when I listened to a talk about time-varying covariates, the speaker asked, rhetorically, what some examples of time-varying covariates are? The fist example, ironically, was time itself. Have you considered a time varying covariate model with a time by treatment interaction?

    ------------------------------
    Stephen Simon
    Independent Statistical Consultant
    P. Mean Consulting
    ------------------------------




  • 5.  RE: Splitting the survival curve

    Posted 09-19-2015 12:26

    Shelley-

    I am relatively new to this forum but did an extensive amount of reliability work in the high tech industry so I thought I should chime in here.  It would help to have a little more info here.  Are there two datasets, one for each product?  If so, I am imaging two survival curves that "cross" each other at some point.  I am a little concerned that you may be reading more into the data post hoc than is justified.  Focussing on the "mid-point" after seeing the data seems a little arbitrary unless there is strong physical knowledge a priori that this might be the mechanism.

    In my experience, I would view this as a problem of fitting and comparing two distributions.  In other words, we want to see if the two products generate the same life distribution.  And if different, how do they differ.   The fact that the curves "cross" might suggest the distributions are different; but it could just be natural variation.

    So I would use JMP (or some other software with strong survival stats apps) to find a life distribution like lognormal or Weibull that fits each dataset reasonably well and then do some tests to see if we can consider the distribution parameters to be the same.  And if not, which are different; for example, shape, scale, location.  This might also help to understand and quanitfy the mechanism.

    Hope this helps!

    -Walt



    ------------------------------
    Walter Flom
    ------------------------------




  • 6.  RE: Splitting the survival curve

    Posted 09-19-2015 12:59
    Hi Shelley,

    Time-varying covariates can occur both naturally and by experimental design. My background is in biomedical research so I'll give a couple examples of each.

    Naturally occurring:
    1. Suppose you are trying to model the trajectory of systolic blood pressure in children through adolescence. The rate of increase may change with onset of puberty in girls. So you'd have an indicator for puberty, that takes effect only at puberty, and might affect both the absolute level of puberty (a jump?) and the rate of change after (slows the rate of increase, starting then.) 

    2. Cognitive performance trajectories in the elderly typically show a gradual decline in performance with age, but a stroke might lead to a drop in function, followed by a slow recovery (if not to baseline), so the "rate of decline" might actually turn into improvement.

    In both cases, you'd need to include in the model a "main effect" for the event (puberty, stroke) that would be time varying, would not click in until the event happened, and then you'd have to include an interaction that would incorporate time since event (would not affect the slope before event.) 

    Experimental: Mainly cross-over designs.
    1. You are treating ovarihysterectomized female mice and measuring bone loss from the time of surgery. Outcome is a bone density measure. The introduction (or cessation) of drug treatment would change the trajectory of bone loss.
    2. You are considering the effect of a community intervention, and one set of communities starts right away and the other set has a delayed introduction (for ethical reasons, you can't have a strict control group that never gets the intervention.) For the delayed-start group, there's a time-varying treatment.
    Hope this helps!

    Laurel A. Beckett, Ph.D.
    Professor and Chief, Division of Biostatistics
    Department of Public Health Sciences
    School of Medicine
    University of California, Davis
    Mailing address:  Division of Biostatistics, MS1C
                                     University of California
                                     One Shields Avenue
                                     Davis, CA  95616
    Phone:  (530) 754-7161
    Fax:       (530) 752-3239





  • 7.  RE: Splitting the survival curve

    Posted 09-19-2015 13:24

    Shelley,

    The standard tests assume stochastic ordering. You have evidence that this is not the case.

    Standard distributions, such as location or scale families, are stochastically ordered.

    However, invoking both location and scale and/or shape, then a model with crossing survival curves can be obtained.

    David

    ------------------------------
    David Bristol
    Statistical Consulting Services
    ------------------------------




  • 8.  RE: Splitting the survival curve

    Posted 09-19-2015 15:43

    You can define time as itself as a time varying covariate with a value of 0 up to time=t and =1 after.  This allows for non proportional hazards of the type you describe.  SAS documentation has examples. Also Google hearing the proportional hazards assumption.

    ------------------------------
    Greg Maislin
    Biomedical Statistical Consulting
    ------------------------------




  • 9.  RE: Splitting the survival curve

    Posted 09-21-2015 15:30


    Thank you for everyone's  comments.  Based on Greg and Simon's input I think I can graphically show the two survival curves have non-proportional hazards.  My primary efficacy test called out in the protocol was the log rank test, nonetheless. 

    But exploratory -wise, I think if I am following the advice, I should carry out a Cox proportional hazard regression analysis with a time-varying covariate/indicator (0 for up to midpoint, and 1 after) and test for significant interaction effect between treatment and indicator?  When I do that both the treatment and treatment-by-indicator effects are significant.  What should the conclusion be then? 


    ------------------------------
    Shelley-Ann Walters
    3M
    ------------------------------




  • 10.  RE: Splitting the survival curve

    Posted 09-21-2015 15:45

    One of the sessions at last week's ASA Biopharmaceutical Section workshop in D.C. addressed this exact problem.  I'm pasting the abstract below, but potentially you could contact the presenters to learn more or they may see this conversation thread themselves.

    PS6a Parallel Session: Statistical considerations of delayed treatment effects in cancer vaccine trials

    09/18/15
    12:45 PM - 2:00 PM
    Thurgood Marshall North

    Organizer(s): Marc Buyse, IDDI Inc.; Jonathan D Norton, MedImmune; Shenghui Tang, FDA; Zhenzhen Xu, CBER/FDA

    Chair(s): Shenghui Tang, FDA

     
     

    In a relatively short period of time, therapeutic cancer vaccines have entered the landscape of cancer therapy. In contrast to the conventional chemotherapeutic drugs, these novel agents stimulate the patient’s own immune response to combat cancer. This indirect mechanism-of-action for vaccines poses the possibility of a delayed onset of clinical effect, due to the time required to mount an effective immune response and the time for that response to be translated into an observable clinical effect. The conventional design and analysis methods based on log-rank test, however, often ignore this delayed effect and result in underestimated sample size with insufficient power, failing to detect the potential effects of the vaccines. More innovative statistical methodologies are needed to address the unique characteristics of therapeutic cancer vaccines in the design and analysis of such a trial. This session will feature speakers from experts in industry, academia and regulatory arenas who will present their research on design and analysis of cancer vaccine trials. Three major topics will be addressed: (1) Sample size calculation considering the delayed treatment effects in cancer vaccine trials; (2) Proper analysis of cancer vaccine trials with delayed treatment effects; (3) Statistical challenges in cancer vaccine trials development from regulatory perspective.

    Sample size and power calculation in therapeutic cancer vaccine trials with delayed treatment effect
    Zhenzhen Xu, CBER/FDA

    Power Calculation for Log-rank Test under a Non-proportional Hazards Model
    Daowen Zhang, North Carolina State University

    Sample Size and Power of Survival Trials in Group Sequential Design with Delayed Treatment Effect
    Jianliang Zhang, MedImmune, LLC



    ------------------------------
    Julie Brevard
    Director of Statistics
    Idera Pharmaceuticals
    ------------------------------




  • 11.  RE: Splitting the survival curve

    Posted 09-21-2015 15:49

    Hi Shelley-Ann:

    Check out the following article: 

    Uno, et al. (2014).  Moving beyond the hazard ratio in quantifying the between-group difference in survival analysis.  J. Clin Oncology, 32(22) 2380- 2385.   

    Some helpful suggestions in there.  

    You have a difficult situation - it might be helpful to look at the patient characteristics of patients who have the event early vs those who have the event late to see how they differ.  Perhaps one therapy is more helpful for the sickest patients but the other therapy better for those less sick.

    ------------------------------
    Roy Tamura
    Associate Professor
    University of South Florida
    ------------------------------




  • 12.  RE: Splitting the survival curve

    Posted 09-21-2015 16:03

    Shelley-

     

    Is there any chance you can share the actual data or a "disguised" version of the data with us so we can make sure we are giving you good advice?

     

    Regards,

     

    -Walt






  • 13.  RE: Splitting the survival curve

    Posted 09-21-2015 16:13

    Shelly,

    The Log-rank test belongs to a family of tests that compares hazard rates.  Different family members use different weight function, many of which are available in SAS PROC LIFETEST.  It is stated on Page 213 of Klein and Moeschberger that 'In most applications of the methods, our strategy is to compute the statistics using the log-rank weights and the Gehan weight. ... In some applications, one of the other weight functions may be more appropriate, based on investigator's desire to emphasize either late or early departures between the hazard rates".  

    Crossed hazard rates could be due to random noise. Emphasizing too much on the crossing point might lead to results of little value.

    Proportional hazard model can be verified by artificially introducing time as a dependent variable and check if its co-efficient is significantly non-zero. However, time should not be a dependent variable in your final model.

    As to time-dependent covariates, they should be used with caution, because time to event is your response variable. Please see Page 307-308 of Klein and Moeschberger for a brief discussion.

           


    ------------------------------
    Qing Kang
    Chief Scientist
    Statistical Intelligence Group, LLC
    ------------------------------




  • 14.  RE: Splitting the survival curve

    Posted 09-21-2015 17:05

    Notwithstanding the benefits of more sophisticated approaches and the benefits of pre-specification,  from an exploratory data analysis point of review, following a highly significant time interaction, it makes sense to me to characterize this in terms of relevant effect sizes., i.e., hazards ratios.   In this case, you can re-estimate your model and determine HR's based on follow-up from time = 0 to time = cut point and then from time = cut point to time = end of last follow-up.   Failures prior to time = cut point are, of course, not included when estimating the second phase of follow-up.

     

     






  • 15.  RE: Splitting the survival curve

    Posted 09-22-2015 15:10

    Well, I would have used a continuous version of time rather than a cut-point, but that's not too critical. Generally, when you find an interaction in a statistical model, you look at some graphs and give a subjective interpretation of the interaction. There are fancy tests to distinguish between a qualitative interaction and a quantitative interaction, but these are not really needed here.

    I'm guessing that your graph shows pretty clearly that one group is at higher risk early and the other group is at higher risk late. Nothing more than that needs to be said, other than a careful warning in the discussion section of the paper that the test of interaction was developed after looking at the data. That makes it largely exploratory in nature and your result needs replication before it can be considered a definitive finding.

    Then ask for a three million dollar research grant to conduct this replication.

    ------------------------------
    Stephen Simon
    Independent Statistical Consultant
    P. Mean Consulting
    ------------------------------