Journal of the National Cancer Institute Advance Access originally published online on February 10, 2009
JNCI Journal of the National Cancer Institute 2009 101(4):216-217; doi:10.1093/jnci/djn515
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
© The Author 2009. Published by Oxford University Press.
EDITORIALS |
Surrogate and Mediating Endpoints: Current Status and Future Directions
Affiliation of author: Public Health Sciences Division, Fred Hutchinson Cancer Research Center, Seattle, WA
Correspondence to: Ross L. Prentice, PhD, Director, Public Health Sciences Division, Fred Hutchinson Cancer Research Center, 1100 Fairview Ave N, M3-A410, PO Box 19024, Seattle, WA 98109-1024 (e-mail: rprentic{at}fhcrc.org).
The identification of surrogate endpoints that can replace "true" endpoints in clinical trials could provide an important advance for the evaluation of therapeutic or preventive interventions. Outcome events that are more frequent in occurrence and more proximate in time, compared with customary disease-specific mortality or incidence outcomes, could give answers that are based on smaller trials of shorter duration. However, reliance on surrogate outcomes is justifiable only if treatment comparisons that are based on a surrogate are a faithful reflection of comparisons that are based on the true endpoint.
I took this perspective, almost 20 years ago (1), in defining a surrogate outcome to be "a response variable for which a test of the null hypothesis of no relationship to the treatment groups under comparison is also a valid test of the corresponding null hypothesis based on the true endpoint," thinking that this was a minimal requirement for a short-term outcome to provide some reliable treatment effect information for the longer term outcome. However, this apparently simple requirement translates to some strong restrictions on the relationship of the treatment to the surrogate and true outcomes: Consider a treatment indicator variable x, a time to response surrogate S, and a time to response true outcome T. A dependence of T on x will imply a dependence of S on x if
- the hazard rate for T depends on S and
- the hazard rate for T given S does not depend on x.
The first criterion is typically readily verified empirically, whereas the second, which requires the surrogate to fully mediate the treatment effect on true outcome, is not. Rather, empirical data alone, even if extensive, will not provide certainty concerning criterion 2. Criterion 2 typically entails detailed knowledge of the biological pathways whereby x may affect T and detailed knowledge about the time course of such effects—knowledge that one would not expect to be available if there is uncertainty concerning whether x has any effect on T. For criterion 2 to hold, the surrogate must be comprehensive in being responsive to all pertinent pathways and the implications of the surrogate event occurrence for the true outcome risk must be equivalent in each treatment group being compared.
The article by Ray et al. (2) in this issue considers both the occurrence of distant metastases and general clinical treatment failure as potential surrogate outcomes for prostate cancer–specific death in the context of evaluating the effect of long-term androgen deprivation therapy among prostate cancer patients with locally advanced disease. Among patients who were alive 3 years after randomization, the hazard ratio for prostate cancer–specific death, when long-term androgen deprivation therapy was compared with control treatment, was 0.76 (95% confidence interval [CI] = 0.51 to 1.11) among patients without distant metastases and 0.95 (95% CI = 0.60 to 1.50) among patients with distant metastases. In the context of a corresponding unconditional hazard ratio of 0.69 (95% CI = 0.52 to 0.93) for prostate cancer–specific mortality, these analyses convey useful information about the importance of a reduction of distant metastasis in mediating the treatment effect on prostate cancer–specific survival, but they do not provide persuasive information concerning the ability of distant metastases to fulfill criterion 2. For example, the estimated prostate cancer–specific mortality rate among men without distant metastases is estimated to be 24% lower in the long-term androgen deprivation group than in the control group. Is this a chance observation or does longer term deprivation have some impact, for example, on local or regional recurrence or on the timing of distant metastasis detection?
Consideration of general clinical treatment failure as a potential surrogate (2) can be viewed as an effort to encompass pathways, in addition to reduction in distant metastases incidence, whereby the treatment may affect prostate cancer–specific death rates. General clinical treatment failure was defined as the time to first occurrence of local prostate recurrence, documented regional or distant metastasis, initiation of androgen deprivation therapy after protocol-directed treatment, or a prostate-specific antigen level of 25 ng/mL or higher after completion of radiation therapy. In spite of the stringency of criteria 1 and 2, a further criterion, as described previously (1), is needed to ensure that a dependence of S on x translates to a dependence of T on x. In this context, one can ask whether some events are included in general clinical treatment failure that help to establish an effect of x on S but have little or no implication concerning an effect of x on T. For example, the initiation of extra-protocol androgen deprivation therapy evidently may differ between treatment groups for purely artifactual reasons (ie, such therapy is unlikely in the first 2 years after randomization for men assigned to the long-term deprivation group). Also, the potential surrogate, general clinical treatment failure, may have a substantial noise component due to issues in defining local recurrence or due to patient concerns leading to unnecessary androgen deprivation therapy. Beyond these issues, the hazard ratio for long-term androgen deprivation vs control treatment was 0.88 (95% CI = 0.48 to 1.63) before general clinical treatment failure and 0.81 (95% CI = 0.58 to 1.14) after general clinical treatment failure, so that once again these empirical data do not provide convincing support for criterion 2. The issue of full mediation by a potential surrogate is of more than academic interest, in view of multiple examples in which a clear treatment effect on a presumed surrogate did not translate to a corresponding effect on the clinical outcome of interest in a subsequent trial (3).
Another surrogacy issue concerns the application of a presumed surrogate to the evaluation of treatments other than that used to argue surrogacy or even to the same treatment in different populations. How similar do the new treatments, or the new populations, need to be to those originally studied? The answer would seem to be very context dependent and again would evidently require knowledge about all relevant biological pathways whereby the new treatment being evaluated may affect the outcome of primary interest.
In an editorial following an earlier article (4) on other potential surrogates for prostate cancer–specific death, Baker (5) made some similar comments and ended with a call for data sharing toward "meta-analytic validation of surrogate endpoint." This is a sensible proposal given the enormous sample sizes that would be required to establish criterion 2 even approximately. In fact, most meta-analytic proposals concerning surrogate outcome evaluation, as summarized previously (6), do not attempt to establish criterion 2 but rather rely on a positive correlation between treatment effect parameter estimates for the surrogate and for the true outcome, to induce a treatment effect estimate for the true outcome in a future trial that is based on data on the surrogate. This approach has some appeal, but it necessarily involves complex modeling of treatment effects jointly on the surrogate and true outcomes. Again, judgments are required as to which previous trials and their treatments have enough commonality with the treatment and population under study to support this type of treatment effect extrapolation. Biological pathways remain important in the meta-analytic context as well, because a surrogate that is not responsive to some aspect of the treatment effect on the outcome of interest can be expected to mislead if this aspect is important for the new treatment.
Given the apparently limited potential to convincingly validate a surrogate outcome, especially for use in the evaluation of a range of future treatments, the focus of treatment effect evaluation on intermediate outcome variables could constructively turn to mediation analyses. As already noted, the article by Ray et al. (2) demonstrates that reduction in distant metastases appears to be a valuable component of the effects of long-term androgen deprivation on prostate cancer–specific mortality. The identification of a set of intermediate outcomes that collectively may be rather comprehensive concerning the pathways whereby a class of treatments could exert beneficial or harmful health effects may be a practical and useful aspect of the research agenda. Use of data of this type in conjunction with data from observational studies to relate such intermediate marker values jointly to the occurrence of various pertinent health outcomes could put researchers in a position to make an informed judgment concerning whether to proceed with a long-term trial having disease incidence or mortality outcomes, including the possibility of early warning concerning adverse effects. The advent of high-dimensional biological data, including data from gene expression and proteomic platforms, makes practical the production of detailed treatment response information from tumor tissue, blood, or other biospecimens. These data, in conjunction with data on changes in established clinical parameters or risk factors, have the potential to provide an unusually comprehensive set of intermediate outcomes, to facilitate the screening of therapeutic or preventive interventions, and to enhance both the monitoring and interpretation of clinical trials. Related methodological and applied research would seem to be timely and important.
REFERENCES
1. Prentice RL. Surrogate endpoints in clinical trials: discussion, definition and operational criteria. Stat Med (1989) 8(4):431–440.[Web of Science][Medline]
2. Ray ME, Bae K, Hussain MHA, Hanks GE, Shipley WU, Sandler HM. Potential surrogate endpoints for prostate cancer survival: analysis of a phase III randomized trial. J Natl Cancer Inst (2009) 101(4):228–236.
3. Fleming TR, DeMets DL. Surrogate endpoints in clinical trials: are we being misled? Ann Intern Med (1996) 125(7):605–613.
4. Petrylak DP, Ankerst DP, Jiang CS, et al. Evaluation of prostate-specific antigen declines for surrogacy in patients treated in SWOG 99-16. J Natl Cancer Inst (2006) 98(8):516–521.
5. Baker SG. Surrogate endpoints: wishful thinking or reality? J Natl Cancer Inst (2006) 9(8):502–503.
6. Burzykowski T, Molenberghs G, Buyse M. The Evaluation of Surrogate Endpoints (2005) New York, NY: Springer Publishing Co. 1–399.
Related Article in JNCI
CiteULike 
Connotea 
Del.icio.us What's this?
J Natl Cancer Inst 2009 101: 215.
This article has been cited by other articles:
|
|
 |
|
  B. M. Psaty and R. L. Prentice Variation in Event Rates in Trials of Patients With Type 2 Diabetes JAMA, October 21, 2009; 302(15): 1698 - 1700. [Full Text] [PDF]
|
|
|
|
 |
|
 A. Latib, J. Cosgrave, and A. Colombo Bimodal distribution of angiographic measures of restenosis: what does it mean? Heart, October 1, 2009; 95(19): 1556 - 1558. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||