Evaluating the Optimal Timing of Angiography
Landmark or off the Mark?
- cardiovascular diseases
- coronary angiography
- data interpretation, statistical
- outcome assessment
The study by Tricoci and colleagues1 in the present issue of Circulation concludes that shortening the time from hospital admission to coronary angiography was associated with fewer ischemic outcomes with no increased bleeding. Although we know that use of an early invasive strategy in patients with non–ST-segment–elevation acute coronary syndromes is associated with improved outcomes, the optimal time to perform coronary angiography in those scheduled to receive an invasive strategy is unknown.
Article p 2669
The Tricoci et al1 study capitalized on the clinical data collected as part of the Superior Yield of the New Strategy of Enoxaparin, Revascularization, and Glycoprotein IIb/IIIa Inhibitors (SYNERGY) trial2 to study this question. Because the goal of the SYNERGY trial was to compare the outcomes of patients treated with enoxaparin versus unfractionated heparin, (1) patients were not randomized to different times to angiography (such as ≤6 hours, 6 to 12 hours, 12 to 18 hours, etc) after hospital arrival, and (2) some patients may have died or experienced an adverse event before receiving angiography. The authors adopted 2 different analytical strategies to address these issues: a landmark analysis3 and an inverse-probability–weighted approach.4 These 2 approaches differ in their basic assumptions and in the populations to which they apply.
The Landmark Method
This method was proposed in the early 1980s in cancer studies in which many researchers had compared survival rates of treated patients whose tumors responded to a therapy to those of treated patients whose tumors did not respond. The erroneous conclusion often made from this comparison was that responders survived longer than nonresponders (which implies that one should increase the percentage of patients who respond). A bias exists because the length of patient survival will affect the likelihood that a patient becomes a responder. This is due to the fact that patients who die earlier in the study will not have a chance for their tumor to respond to the treatment and will therefore make the nonresponder group have worse survival.
As a fix to this problem, the landmark method, in which the investigator selects a fixed time (“the landmark”) as initiation of therapy, was proposed to analyze such data. It is important to note that in this setting, randomization ensures comparability between patients assigned to treatment groups; outcome differences between responders and nonresponders could be related to pretreatment patient characteristics but not to patient or physician choices. Patients are followed forward in time from the landmark to determine if survival from the landmark depends on the patient’s status at the landmark—thus, the clock is reset at the landmark. A single landmark is selected before analysis. Patients who died or went off protocol before the time of the landmark are excluded from the analysis.
It is important to assess how this framework applies to the Tricoci study1 (see the Figure). Consider the evaluation of angiography within 6 hours of hospital arrival. The investigators compared myocardial infarction and mortality events 30 days from randomization into the SYNERGY trial between patients who had an angiography within 6 hours of hospital arrival and patients who underwent coronary angiography later or not at all. Patients who had a myocardial infarction or who died within 6 hours of the angiography were excluded. First, many patient and physician factors may affect the decision about who gets early versus late angiography. This is in contrast to the original assumptions of a landmark analysis, where responder status is not affected by patient or physician choice. Second, the outcomes (myocardial infarction and death) should be evaluated from the landmark and not from randomization. This is not a real problem if patients are randomized at hospital presentation, given that the landmark is basically measured at time of randomization, eg, 6 hours from hospital arrival. However, not resetting the clock to the landmark time poses more problems when using 3 days as the landmark. Third, the conclusions from this analysis are conditional on angiography status in the subset of patients who are alive and free of myocardial infarction at 6 hours after hospital arrival. One could imagine many decision-makers and factors that affect who gets coronary angiography. Therefore, the results apply to a highly selected population from which it may be difficult to generalize.
The Inverse-Probability–Weighted Method
This approach uses methods proposed for studies in which the level of treatment is tailored through time to an individual’s changing health status.5 The idea is to model treatment time as a random variable that depends on patient characteristics and the treating physician’s characteristics while accounting for treatment-initiating or censored events. The approach involves first modeling the probability of receiving treatment at a particular time, accounting for any treatment-censoring events, and then weighting the outcomes by the inverse of the estimated probabilities. Although there are several technical details, the key assumption is that, once the patient’s clinical history is known up to a particular time, the decision to initiate the therapy at that particular time does not depend on future prognostic factors—the “no unmeasured confounders” assumption.
In contrast to the landmark approach, treating time to angiography as a random variable rests on more plausible assumptions and utilizes the full study population. The censoring events are the same as those the authors identify in the landmark method, such as death. Tricoci et al1 estimated a Cox model to predict time to angiography, accounting for censored events, and then computed the weighted mean ischemic event rate. The assumption of no unmeasured confounders cannot be verified directly but could be bolstered through sensitivity analyses. This analytical strategy holds much promise, especially when we can add supplemental information about confounders.
With increasing interest in using clinical data to assess important policy decisions, it will be important that investigators, reviewers, and readers carefully assess both the assumptions made and their plausibility. The inverse-probability–weighted method is an elegant approach that makes clear assumptions and is potentially generalizable. The assumptions for a landmark analysis are far more suspect and often just off the mark.
The opinions expressed in this article are not necessarily those of the editors or of the American Heart Association.
Tricoci P, Lokhnygina Y, Berdan LG, Steinhubl SR, Gulba DC, White HD, Kleiman NS, Aylward PE, Langer A, Califf RM, Ferguson JJ, Antman EM, Newby LK, Harrington RA, Goodman SG, Mahaffey KW. Time to coronary angiography and outcomes among patients with high-risk non–ST-segment–elevation acute coronary syndromes: results from the SYNERGY trial. Circulation. 2007; 116: 2669–2677.
Anderson JR, Cain KC, Gelber RD. Analysis of survival by tumor response. J Clin Oncol. 1983; 1: 710–719.
Johnson BA, Tsiatis AA. Semiparametric inference in observational duration–response studies, with duration possibly right-censored. Biometrika. 2005; 92: 605–618.